Quantity implicatures
Bart Geurts
December 21, 2009. To appear with Cambridge University Press.
Table of contents
Preface
1
Introduction
2
1 Grice
1.1 Saying vs. implicating . . . . . . . .
1.2 Discourse and cooperation . . . . . .
1.3 Conversational implicatures . . . . .
1.4 Generalised vs. particularised . . . .
1.5 Cancellability . . . . . . . . . . . . .
1.6 Gricean reasoning and the pragmatics
2 The
2.1
2.2
2.3
2.4
2.5
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
of what is
Standard Recipe for Q-implicatures
The Standard Recipe . . . . . . . . . . .
Inference to the best explanation . . . .
Weak implicatures and competence . . .
Relevance . . . . . . . . . . . . . . . . .
Conclusion . . . . . . . . . . . . . . . . .
. . .
. . .
. . .
. . .
. . .
said
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
6
6
11
13
17
19
22
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
27
27
33
37
42
48
3 Scalar implicatures
3.1 Horn scales and the generative view . . . . . . . . .
3.2 Implicatures and downward entailing environments
3.3 Disjunction: exclusivity and ignorance . . . . . . .
3.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
49
50
54
59
64
4 Psychological plausibility
4.1 Charges of psychological inadequacy
4.2 Logical complexity . . . . . . . . . .
4.3 Abduction . . . . . . . . . . . . . . .
4.4 Incremental processing . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
66
67
69
70
71
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
4.5
4.6
4.7
The intentional stance . . . . . . . . . . . . . . . . . . . . . .
Alternatives . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5 Nonce inferences or defaults?
5.1 True defaults . . . . . . . .
5.2 Strong defaultism . . . . . .
5.3 Weak defaultism . . . . . .
5.4 Contextualism . . . . . . . .
5.5 Conclusion . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
6 Intentions, alternatives, and free choice
6.1 Free choice . . . . . . . . . . . . . . . . .
6.2 Problems with the Standard Recipe . . .
6.3 Intentions first . . . . . . . . . . . . . . .
6.4 Free choice explained . . . . . . . . . . .
6.5 Comparing alternatives . . . . . . . . . .
6.6 Two flavours of Q-implicature . . . . . .
6.7 Conclusion . . . . . . . . . . . . . . . . .
7 Embedded implicatures: the problems
7.1 The problems . . . . . . . . . . . . . .
7.2 Varieties of conventionalism . . . . . .
7.3 Against conventionalism . . . . . . . .
7.4 Conclusion . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
8 Embedded implicatures: a Gricean approach
8.1 Disjunction . . . . . . . . . . . . . . . . . . .
8.2 Belief reports . . . . . . . . . . . . . . . . . .
8.3 Factives and other presupposition inducers . .
8.4 Indefinites . . . . . . . . . . . . . . . . . . . .
8.5 Contrastive construals and lexical pragmatics
8.6 Conclusion . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
73
76
80
.
.
.
.
.
82
84
86
92
98
102
.
.
.
.
.
.
.
103
105
108
110
115
119
123
129
.
.
.
.
131
133
142
151
160
.
.
.
.
.
.
162
163
165
168
173
178
186
Afterword
189
Notation and abbreviations
192
Bibliography
194
Preface
Quantity implicatures had been haunting me for about a decade, when I finally
decided it was time to put my mind at rest. Writing this book proved to be
of considerable therapeutic value, but of course my hope is that it will be
of value (therapeutic or otherwise) to its readers as well. Since I have been
thinking about this topic for quite a few years, there was ample opportunity
for changing my mind, which I did a number of times. I’ve had my heretical
phases (most recently in my 2005 paper on disjunction), but they are behind
me now, and I’m glad to be back in the Gricean fold.
I have tried to write a text that is accessible to a fairly broad audience,
including students of linguistics, philosophy, and psychology. In recent years,
the literature on my topic has become increasingly esoteric. I don’t welcome
this development, and I don’t see it was inevitable, either. Be that as it
may, my aim was to write a book that presupposes little more than general
intelligence and common sense, though some familiarity with a handful of
very basic linguistic and logical notions may prove to be useful. (A brief
compendium of the latter will be found at the end of the book.)
For various kinds of help and support at various stages along the way,
I would like to thank Patrick Blackburn, Richard Breheny, Vincent Homer,
Janneke Huitink, Hans Kamp, Napoleon Katsos, Rick Nouwen, Ira Noveck,
François Récanati, Philippe Schlenker, Katrin Schulz, Benjamin Spector, Dan
Sperber, Rob van der Sandt, Tanja van Hummel, Noor van Leusen, Robert
van Rooij, Bob van Tiel, and an anonymous reviewer. Special thanks is due
to Corien Bary, Anne Hökberg, and Larry Horn, for their profuse comments
on the first complete draft; to Emmanuel Chemla, who kindly demolished the
first version of Chapter 6 for me; and to Nausicaa Pouscoulous, for joining
me in my chase of embedded hippos. Finally, I’m grateful to the Netherlands
Organisation for Scientific Research (NWO) for financial support in the last
throes of the project.
Introduction
This book is all about one simple idea: that speakers convey information not
only by what they say, but also by what they don’t say. In fact, the idea is
so obvious that we may never know who had it first. What is known is that
around the mid-19th century, John Stuart Mill thought it obvious enough to
be mentioned almost in passing:
If I say to any one, “I saw some of your children to-day”, he might be justified
in inferring that I did not see them all, not because the words mean it, but
because, if I had seen them all, it is most likely that I should have said so:
though even this cannot be presumed unless it is presupposed that I must
have known whether the children I saw were all or not. (Mill 1865: 442, also
cited by Horn 2009)
One century later, it was H. Paul Grice who saw that this simple idea contains
the gist of a general framework for pragmatics, based on the premiss that
discourse is a joint project undertaken by speakers who expect each other to be
cooperative. It is this expectation, according to Grice, which gives rise to the
pragmatic inferences he calls “conversational implicatures”. The inference Mill
refers to is such an implicature; more specifically, it is a quantity implicature.
If it is so simple, why does it take a book to explain? There are several
reasons. To begin with, it turns out that quantity implicatures are somewhat
more complex than meets the eye (Chapter 2), but that is not the main reason. More important is that the extensive literature on quantity implicatures
contains a number of distractors that we will have to clear out of the way.
One is that the bulk of that literature has focused its attention on one species
of quantity implicatures, i.e. scalar implicatures, which has led to a distorted
view on the genus (Chapter 3). Mill’s example is a case in point. If a speaker
says,
3
(1)
I saw some of your children today.
we can generate an alternative statement he could have made simply by supplanting “some” in (1) with a stronger expression of about the same complexity:
(2)
I saw all of your children today.
Then the implicature is that, for all the speaker knows, this alternative statement doesn’t hold. There has been a tendency in the literature to treat all
quantity implicatures on this model, but the trouble is that it isn’t general
enough. For instance, it already fails in simple cases like the following:
(3)
A: Who is coming to the meeting?
B: Albinoni, Boccherini, and Corelli.
B’s answer may well implicate that apart from Albinoni, Boccherini, and
Corelli, nobody else is coming to the meeting, and if this is inferred it must
a quantity implicature. However, it is evident that this inference cannot be
obtained by replacing an expression in (3B) by one that is stronger and about
equally complex, and then denying the result.
What is needed, I will argue, is a shift of perspective that may seem suspiciously subtle at first, but is entirely in the Gricean spirit and a prerequisite
for a unified theory of quantity implicatures (Chapter 6). Instead of asking
why the speaker didn’t say (2), the question should be whether the speaker
might believe that (2) is true. In this case, the answer is likely to be “no”,
for if the speaker believed that (2) is true, he would have uttered this sentence rather than (1), and thus we arrive at the conclusion that the speaker
does not believe (2). I call this the “intention-based” approach. Unlike the
standard scheme, it carries over to examples like (3) without a hitch. For A
may now reason as follows: “Could it be that, according to B, more than three
people are coming to meeting? Presumably not, because then B would have
said so. Hence, I’m entitled to conclude that, for all B knows, only Albinoni,
Boccherini, and Corelli are coming to the meeting.” Furthermore, this line of
explanation extends to free choice inferences, like the following:
(4)
Julius may be in Amsterdam or Berlin.
; Julius may be in Amsterdam.
; Julius may be in Berlin.
Despite the fact that they are obviously licit in some sense, these inferences
have proved hard to account for. As I will argue in Chapter 6, one of the main
attractions of the intention-based approach is that it allows us to explain them
as run-of-the-mill quantity implicatures.
4
Introduction
Another distracting factor we will have to engage with is related to the first.
It is the notion that some quantity implicatures hold by default (Chapter 5).
Again, scalar implicatures are the prime examples. It is tempting to suppose
that an utterance of (1) will normally give rise to the inference that, for all the
speaker knows, (2) isn’t true. Apparently, this intuition is so compelling that
many scholars have accepted it at face value. Nonetheless, I will argue that
it is mistaken, and has done more than its share to keep implicature theorists
off the right track.
Perhaps the main challenge for a Gricean theory of quantity implicatures,
and a further reason why it takes more than thirty pages to develop, is that
every once in a while it appears as if a quantity implicature occurs within the
scope of an attitude verb, for example:
(5)
Cleo believes that some of her shoes have been stolen.
This may be understood as implying that Cleo believes that not all her shoes
were stolen, and this interpretation seems to require that a “not all” inference
is drawn within the scope of “believes”. If that is the case, the inference in
question cannot be a quantity implicature, because quantity implicatures, like
all conversational implicatures, are derived on the basis of a speech act made
by the speaker, and speech acts are made by way of complete sentences. Hence,
there is no way a bona fide implicature could be derived within the scope of
any expression.
This is commonly known as the problem of “embedded implicatures” (which
is strictly speaking a contradiction in terms, since by their very nature implicatures cannot be embedded). Embedded implicatures pose a problem for
Gricean pragmatics in two ways. On the one hand, they have prompted a
rash of proposals for analysing quantity implicatures, and especially scalar implicatures, with non-pragmatic methods. For instance, it has been suggested
that “some” actually means “some but not all”, and if this were true, examples
like (5) wouldn’t be much of a problem. There is quite a variety of such conventionalist proposals, which are to be discussed and criticised in Chapter 7.
On the other hand, even if this type of approach is on the wrong track, that
doesn’t give us a pragmatic solution to the problem of embedded implicatures.
Chapter 8 takes up the challenge by arguing that there isn’t one big problem
but rather several smaller ones, most of which can be solved by showing that
there are independent factors that interact with standard Gricean reasoning
in such a way that embedded implicatures seem to arise. Put otherwise, I will
argue that, by and large, embedded implicatures are a mirage. Still, there
5
are exceptional cases in which, e.g., “some” does mean “some but not all”, and
these cannot be construed as conversational implicatures. However, I will try
to show that they, too, can be accommodated within the framework of Gricean
pragmatics.
In the meantime, it will have become evident that this is an essay in pure
pragmatics; Gricean pragmatics, to be exact. However, although the theory
to be presented in the following pages is squarely in the spirit of Grice’s philosophy, it does not slavishly adhere to the letter of his writings. There are
three points, in particular, in which I will deviate from the official Gricean
party line. To begin with, whereas Grice usually (though by no means always)
studied discourse from the speaker’s perspective, I will resolutely adopt the
hearer’s point of view: my main interest lies in the interpretation of linguistic
utterances. As a consequence, I will use some of the standard Gricean terminology in a way that is strictly speaking deviant (though it may be noted
that this sort of terminological abuse is rife in the literature, even if it usually
goes unacknowledged). In particular, this holds for the term “implicature”,
which Grice uses for certain commitments the speaker incurs by virtue of his
utterances, whereas I will treat implicatures as hearers’ inferences about such
commitments. This is merely a difference in perspective, not opinion.
The second point is related to the first. In my view, a theory of interpretation—any theory of interpretation—should be of a piece with a psychological
theory of language processing. This is not to say that a theory of interpretation
should be a processing theory, but it is to say that pragmatics and psychology
should mesh and that experimental data can and must be brought to bear
on theories of interpretation. In this respect, too, I deviate from Grice, who
(at least in his writings) never showed much of an interest in experimental
psychology.
Thirdly, and most importantly, there is the much-debated issue of “what
is said”: Grice’s term for what is also known as the “proposition”, “literal
meaning”, or “truth-conditional content” carried by an utterance. Since conversational implicatures are calculated on the basis of what is said, it is of
considerable importance to establish what is said by any given utterance and
it is especially important to determine if and how pragmatic factors may affect
truth-conditional content. It is widely agreed that Grice’s own views on the
matter were not entirely satisfactory, and I will go with the general consensus
in this point, and argue, moreover, that truth-conditional content is dependent on pragmatic reasoning to a much larger extent than Grice was willing to
allow for. But nevertheless, I will also argue, against the consensus, that this
amendment is entirely consistent with the spirit of Grice’s general programme.
Grice
1
The concept of conversational implicature was developed by Grice in his groundbreaking work of the 1960s, which culminated in the William James Lectures
he delivered at Harvard University in 1967. These lectures had a formidable
impact from the very start, but were slow to appear in print, the key installments for our topic being “Logic and conversation” (1975) and “Further notes
on logic and conversation” (1978). It wasn’t until 1989 that the William James
Lectures were published collectively, as part of a selection of Grice’s papers,
“Studies in the way of words”.
The theory of quantity implicatures that will unfold in the following chapters is thoroughly Gricean in spirit, although occasional departures from the
letter of Grice’s writings will be hard to avoid. In this chapter, I will survey Grice’s views on conversational implicatures, and mark the chief points at
which I deviate from his opinion. The discussion will be confined to matters
that are relevant to the purposes of this book, which is to say that important
parts of Grice’s work will be treated lightly, if that. For a more comprehensive
introduction to the Gricean philosophy of language, I refer to Neale’s (1992)
excellent review article on “Studies in the way of words”.
1.1 Saying vs. implicating
Consider the following sentences:
(1)
a. Harry is rich but dull.
b. Harry is rich and dull.
Despite the undeniable fact that there is considerable semantic overlap between
these sentences, it would seem that (1a) has a richer meaning than (1b). Both
1.1 Saying vs. implicating
7
sentences entail that Harry is rich as well as dull, but while this is all that is
conveyed by (1b), (1a) conveys a bit more, viz. that one way or another there
is a contrast between being rich and being dull.1 Hence, we are inclined to say,
pre-theoretically, that (1a) and (1b) have different meanings. More generally,
in everyday practice we are liable to conclude that two sentences have different
meanings whenever we observe that they don’t communicate exactly the same
message. One of the the most important insights to have emerged from the
modern philosophy of language is that this practice is mistaken: our pretheoretical notion of meaning covers an amalgam of very different species of
content, which must be differentiated phenomenologically (they don’t behave
alike) as well as etiologically (they have different causes). Consequently, one of
the main concerns of theories of interpretation is to isolate species of content
and study their properties as well as the ways in which they interact. Small
wonder, therefore, that in semantics and pragmatics demarcation issues are
the order of the day. One theory’s entailment is another’s presupposition or
implicature.
Consider again the contrast between (1a) and (1b). According to Frege
(1892, 1918) both sentences have the same meaning; they express the same
proposition. Meaning, in Frege’s book, is what determines whether or not
a sentence is true, in a given context, and this is the same in either case:
Frege considers both sentences true if Harry is rich and dull; otherwise they
are both false. Frege concedes that (1a) “intimates” a contrast between the
property of being rich and that of being dull, but he denies that this is part of
the sentence’s truth conditions. Therefore, Frege would say that (2b), while
surely infelicitous, does not have contradictory truth conditions, as (2a) does:
(2)
a. *Harry is rich but dull, but he isn’t rich.
b. ?Harry is rich but dull, though I wouldn’t want to suggest that there
is a contrast between these two properties.
If the contrast between (2a) and (2b) seems a bit fragile, there are other ways
of bringing out the difference between truth-conditional and intimated content.
Consider, for example, the following pair of discourses:
(3)
a. Harry is rich and dull, and that’s a good thing.
b. Harry is rich but dull, and that’s a good thing.
In both cases, the intended referent of “that” is the fact that Harry is rich and
dull: the second conjunct of (3b) does not imply that it is a good thing that
1.
This is a bit of a simplification, as pointed out by Dummett (1973), for example.
8
Chapter 1
there is contrast between being rich and being dull. Apparently, the demonstrative pronoun ignores the intimated content of the first conjunct, which
supports the view that truth-conditional and intimated content are distinct.
The following example makes the same point in a different way:
(4)
a. Most of these girls are rich and dull.
b. Most of these girls are rich but dull.
(4a) says that for most x in a given set of girls G it is the case that (i ) x is rich
and (ii ) x is dull. That is, the entire content of “rich and dull” is attributed
to each of a majority in G. By contrast, (4b) does not say that for most x in
G it is the case that (i ) x is rich, (ii ) x is dull, and (iii ) there is a contrast
between x being rich and x being dull. Rather, what this sentence expresses
is that there is a contrast between being rich and being dull and that most
of the girls are rich and dull. Hence, the intimated content associated with
“but” does not behave as ordinary truth-conditional content would: it seems
oblivious to the fact that it is within the scope of a quantifier.
Frege’s distinction between truth-conditional and intimated content foreshadows Grice’s distinction between “what is said” and “conventional implicature” (Horn 2007). What is said is truth-conditional content; conventional
implicatures have no bearing on the truth value of a sentence. However, in
another way these two notions are closely related, for in Grice’s view both
are conventional in nature—or very nearly so. The contrasts in (1)-(4) are
due to the fact that “but” comes with an implicature that “and” lacks, and
since it is part of the conventions of English that “but” is used this way, Grice
calls it a “conventional implicature”. To a first approximation, at least, what
is said is conventional in much the same way: linguistic conventions assign
truth-conditional meanings to words and grammatical constructions, and the
meaning of a complex expression is determined by the meanings of its parts
and the grammatical constructions they enter into; thus the truth-conditional
meaning of a sentence, “what it says”, is conventional. Grice is aware of the fact
that this picture is somewhat simplified, because truth-conditional meaning is
affected by various contextual factors, and therefore not purely conventional.
This context dependence raises some rather deep issues, which we will dwell on
later in this chapter (§1.6). In the meantime, let’s ignore these complications
and agree that what is said is primarily determined by linguistic conventions.
. . . A digression on truth . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Like Frege, Grice considers conventional implicatures to be truth-conditionally
inert. (In fact, “implicature” is Grice’s cover term for non-truth-conditional
1.1 Saying vs. implicating
9
content.) Considerable quantities of ink have been spilled on the question of
whether or not this is right. To illustrate what the issue is about, suppose I
utter the following:
(5) ?Harry is British but he speaks English.
In view of the well-known fact that English is spoken by the overwhelming
majority of British citizens, there would not normally be a contrast between
being British and speaking English, and therefore my utterance is odd. Now,
the mooted question is this: When I said (5), did I speak falsely or was my
utterance merely defective in some non-alethic sense of the word? In my
opinion, for all the attention it has received, this question is not particularly
interesting, for the following reason. Nobody will deny that my utterance of
(5) carried the information that there is a contrast, and that this information
is false because it fails to agree with the facts. I take it that this much is
uncontroversial. It is also agreed upon, I believe, that a lack of contrast renders
my utterance defective. The question is just whether its defect should be called
“falsity” or something else, and that is a terminological issue.
Whether or not it is merely about terminology, it should be noted that
speakers’ intuitions will not help to resolve the issue. If a professional philosopher were to use words like “true” and “false” as loosely as they are used in
real life, his job would be on the line. The predicate “true” is generally used
for expressing agreement, and not merely to indicate that a claim is factually
correct. Conversely, calling a claim “not true” is a perfectly acceptable way of
saying you disagree, even if there can hardly be a fact of the matter. Here are
some examples from the internet:
(6)
a. Somebody was saying I am a funny and sweet girl. That’s true.
b. The churros with chocolate were lovely, it’s true.
c. It is easier to bottle-feed than to breast-feed. Not true!
In all likelihood, it is a matter of taste whether the speaker of (6a) is a funny
and sweet girl, whether the churros were lovely, and whether bottle-feeding is
easier than breast-feeding. Nonetheless the predicate “true” is freely used in
these cases for expressing authors’ agreement or lack thereof. The upshot of
these observations is that “truth-conditional content” is a theoretical notion
which need not always align with speakers’ intuitions about truth.
The question of whether conventional implicatures have any bearing on
truth is of much less interest than some scholars seem to believe. However,
I suspect that one of the reasons why this question has been discussed so
10
Chapter 1
zealously is that it has been mixed up with another one, viz. whether or not
conventional implicatures should be distinguished from other types of content,
specifically from what is said. That question is important, and can be answered in the affirmative, e.g., on the grounds that conventional implicatures
interact in their own special way with pronouns and quantifiers, as we saw in
our discussion of examples (3) and (4). . . . . . . . . . . . . . End of digression . . . .
Whereas what is said and conventional implicatures are both conventional
(or largely so), conversational implicatures are something different altogether
(and it is unfortunate that Grice chose to use the same term, “implicature”,
for two notions that are so far apart). Conversational implicatures are, first
and foremost, non-conventional ; they are not due to linguistic conventions of
any kind. To explain how such inferences might arise, consider the following
example. Grice (1975/1989: 32) invites us to imagine a person A “standing by
an obviously immobilized car”,2 addressing B as follows:
(7)
A: I am out of petrol.
B: There is a garage round the corner.
Grice observes that B’s utterance allows A to infer that, as far as B knows,
the garage in question is open. It will be clear that this is neither said nor
conventionally implicated: the conventions of the English language aren’t such
that (7B) encodes in any way that the garage is open. I take it that this much is
obvious, but to underscore this point, note that B could have said (8) without
fear of delivering an infelicitous statement:
(8)
There is a garage round the corner, but I’m afraid it is closed.
If it was part of the conventional content of (7B) that the garage is open,
(8) should be infelicitous, which it isn’t. How, then, does the inference come
about? Grice’s answer is really very simple: A assumes that B is trying to
be helpful, and reasons that B’s utterance wouldn’t be helpful if he knew or
suspected that the garage isn’t open. Hence, A concludes, for all B knows the
garage is open. Put otherwise, supposing that B is trying to be cooperative, it
wouldn’t make sense for him to utter (7B) unless he believed the garage was
open, and therefore he probably believes that it is open.
The inference that for all B knows, the garage is open, Grice calls a “conversational implicature”. It arises on the assumption that the speaker is doing
2.
I’ve always been puzzled by Grice’s suggestion that one can see that a car with an empty
tank is “obviously immobilized”.
1.2 Discourse and cooperation
11
his best to make a useful contribution to the discourse. Like their conventional
namesakes, conversational implicatures are truth-conditionally inert, but apart
from that the two types of inference couldn’t be more different, the most important difference being that conversational implicatures aren’t encoded in the
utterances that give rise to them. We will have much more to say about this,
but first let us consider how Grice unpacks the notion that utterances can be
more or less useful.
1.2 Discourse and cooperation
Grice sees a discourse as a collaborative effort.3 A discourse is a joint project
in which the interlocutors aim at achieving one or more common goals. In (7),
the goal is to provide A with petrol. (True, in the first instance, this is A’s
goal, but B’s response shows that he is making A’s cause his own.) Obviously,
there are indefinitely many goals a discourse may serve. People talk with each
other to exchange information, to negotiate a deal, to settle disputes, and so
on.4 But in any given case the number of discourse goals will be fairly small,
and apparent to all interlocutors; they are, after all, common goals.
Given that a discourse is a joint project between interlocutors, it is natural
enough to suppose that an utterance ϕ will normally be interpreted in the
light of the current discourse goals, on the assumption that ϕ was designed by
the speaker to further these goals. In other words, the hearer will assume that
the speaker intends to abide by the
Cooperative Principle
Make your conversational contribution such as is required, at the stage
at which it occurs, by the accepted purpose or direction of the talk
exchange in which you are engaged. (Grice 1975/1989: 26)
This is Grice’s “overarching principle” for an informal theory of cooperative
linguistic behaviour. Grice presents that theory as a collection of more specific
“maxims”, loosely arranged into four rubrics:
3.
Following common usage in pragmatics, I will generally employ the term “discourse” instead
of Grice’s “conversation”.
4. I’m reminded by Larry Horn that language is also used in projects that are non-cooperative
in one way or another: courtroom arguments, political debates, matrimonial fights, and so
on. Note, however, that these are still joint projects which are cooperative at least in the
sense that an exchange of views is supposed to take place.
12
Chapter 1
• Quantity
1. Make your contribution as informative as is required (for the current
purposes of the exchange).
2. Do not make your contribution more informative than is required.
• Quality: Try to make your contribution one that is true:
1. Do not say what you believe to be false.
2. Do not say that for which you lack adequate evidence.
• Relation: Be relevant.
• Manner: Be perspicuous:
1.
2.
3.
4.
Avoid obscurity of expression.
Avoid ambiguity.
Be brief (avoid unnecessary prolixity).
Be orderly.
It bears emphasising that these maxims are not to be thought of as based on
convention, let alone linguistic convention. Rather, Grice sees them reasonable
expectations flowing naturally from a universal theory of cooperative conduct,
of which a theory of cooperative linguistic conduct is merely a special case.
True, Grice never got round to working out such a theory, but the very banality
of most of the maxims suggests rather strongly that his general conception is
a feasible one. Surely, if a speaker wants to be cooperative, it would be a good
idea for him to make his utterances sufficiently informative, speak the truth,
and so on, and surely it is reasonable for hearers to expect speakers to behave
in conformity with such patently sensible rules. In brief, the banality of the
maxims strongly speaks in their favour.
It will be evident that Grice’s maxims are a mixed bag. To begin with, the
Quality maxims, especially the first one, are arguably more important than
the others. Mendacity makes a speaker unreliable; verbosity, obscurity, or
lack of orderliness are just a nuisance. Therefore, Grice suggests, it might be
reasonable to accord a special status at least to the first Quality maxim: other
maxims come into play only on the assumption that this maxim is satisfied.
But the same goes for the maxim of Relation. After all, a speaker who makes
no attempt at making his utterances relevant to the purpose of the discourse
is every bit as uncooperative as one who is mendacious, and more obviously
so.5
5.
Cf. Sperber and Wilson (1995), who argue that Relevance (which is not the same Grice’s
maxim of Relation) is the most important, and indeed the only, maxim.
1.3 Conversational implicatures
13
Concerning the Quantity maxims, Grice observes that the second one
should perhaps be subsumed under the maxim of Relation: to the extent that
an utterance is overinformative, it will be irrelevant to the discourse goals.
Again, the same holds, mutatis mutandis, for the first maxim of Quantity.
Therefore, it will not come as a surprise that many authors have tried to
reduce Grice’s maxims to a leaner set while of course criticising competing
reductionist attempts (see, e.g., Atlas and Levinson 1981, Horn 1984, Sperber
and Wilson 1995, and Levinson 2000).
It seems, then, that every maxim is special in one respect or another, and
Grice’s own classificatory scheme is likely to be redundant. Grice is clearly
aware that this is so. In fact, he appears to take his own proposal with a
considerable grain of salt. For one thing, as we have seen, Grice is at some
pains to point out the defects in the classification he proposes. For another,
the mere fact that he tries to shoehorn his maxims into a taxonomy devised
by the philosopher Immanuel Kant for an entirely unrelated purpose suggests
that Grice considers his proposal tentative at best. All of which is to say that
we shouldn’t set too much store by the precise form in which Grice casts his
theory of discourse. The gist of Grice’s theory is just that interlocutors expect
each other to be cooperative and that this general expectation manifests itself
in many different ways: we expect each other to be striving for relevance and
truth, etc., and these expectations serve to make inferences about the speaker’s
intentional state: what she believes, wants, and so on. Any theory that accepts
this general picture can be called “Gricean”.
1.3 Conversational implicatures
A conversational implicature is “what has to be supposed in order to preserve the supposition that the Cooperative Principle is being observed” (Grice
1975/1989: 39-40). ψ is a conversational implicature licensed by the speaker’s
utterance of ϕ if (i ) ψ is not conventionally associated with ϕ and (ii ) the
hearer has good reasons for believing that, assuming the speaker intends to be
cooperative, he wouldn’t have uttered ϕ unless ψ was the case. Let us have a
look at some of examples given by Grice himself to illustrate how the maxims
may be exploited for inducing conversational implicatures, starting with the
petrol example of §1.1:
(9)
A: I am out of petrol.
B: There is a garage round the corner.
14
Chapter 1
Evidently, the main goal of this mini-discourse is to help A obtain petrol for
his car, and B’s utterance wouldn’t be relevant to this goal unless he believed
the garage was open (or at least might be open), had petrol to sell, and so on.
Hence, the implicature that B believes so is driven by the maxim of Relation.
The next example involves the Manner maxim:
(10) Miss X produced a series of sounds that corresponded closely with the
score of “Home sweet home”.
Here, the speaker goes of his way to avoid the verb “sing”, using a considerable
more prolix expression in its stead. He wouldn’t do so unless he believed, and
wanted to convey to his audience, that the predicate “sing” wasn’t entirely
appropriate, presumably because Miss X’s singing wasn’t very good.
Our third example involves the protagonist of this book, the first maxim
of Quantity:
(11) A: Where does C live?
B: Somewhere in the South of France.
Grice glosses this case as follows:
[B’s answer is] less informative than is required to meet A’s needs. This
infringement of the first maxim of Quantity can be explained only by the
supposition that B is aware that to be more informative would be to say
something that infringed the second maxim of Quality, “Don’t say what you
lack adequate evidence for”, so B implicates that he does not know in which
town C lives. (Grice 1975/1989: 33)
The purpose of the exchange in (11) is that A should learn where C lives. A’s
question is a bid to establish this as the discourse goal, and B’s response signals
his acceptance. Grice’s gloss states rather categorically that B’s infringement
of the Quantity maxim can be explained only by assuming that B lacks the
evidence that would warrant a more specific claim, but that is clearly too
strong. It could be, for example, that B considers his answer precise enough
for A’s purposes (Grice simply states that this isn’t the case, but it might
be), or A and B might be playing a guessing game in which the players aren’t
supposed to give fully specific answers, and there are many more possible
situations in which (11B) would not implicate that the speaker doesn’t know
where C lives, though it may be that such situations tend to be rather special.
It is always possible to invent scenarios in which an implicature that seems
natural and even compelling at first would fail to arise. But this is just to say
that people don’t arrive at implicatures by rigidly applying a set procedure,
1.3 Conversational implicatures
15
and despite his momentary lapse in the cited passage, Grice is perfectly aware
that this is not how things go. A conversational implicature is more like a stab
at explaining a linguistic act, and the logic of this sort of stabbing (if indeed
it has a logic) is quite unlike the classical logics in which valid conclusions are
necessitated by their premisses. We will revisit this theme in the following
chapter (§2.2).
Let me now give a somewhat rough general format for the derivation of
conversational implicatures:
i. It starts with a speaker S performing a particular kind of action: S utters
a sentence ϕ.
ii. In view of the current discourse goals, S’s utterance wouldn’t have been
in accordance with the Cooperation Principle unless ψ was the case.
iii. Hence, ψ must be the case.
The implicature, i.e. ψ, will generally be of the form “S believes that ...”,
“S doesn’t know whether ...”, etc. But although ψ will often be about the
speaker’s knowledge state, it may be about other types of intentional state, too.
For example, in a restaurant the statement “We haven’t ordered yet” may be
used to implicate that the speaker wants the waiter to take his order. Similarly,
though ϕ will usually be a statement (Grice only discusses statements, and
the formulation of some of the maxims is tailored to this type of speech act),
it may also be a question or command, for example:6
(12) Please give me a pen or pencil.
This will usually imply that the addressee need not provide both a pen and
a pencil, and this inference is naturally explained as a quantity implicature:
if the request had been for a pen and a pencil, the speaker should have said,
“Please give me a pen and a pencil”, and since she chose to say (12) instead,
that’s probably not what she meant.
It is important to note that it is only speech acts that give rise to implicatures, where a speech act may be defined, somewhat loosely, as the utterance
of a sentence that commits the speaker to having a certain belief, desire, etc.
Since speech acts are not sentences, but rather actions of a specific kind, it
would be a logical faux pas to say that this or that sentence carries such-and6.
Here I’m assuming that the exclusive interpretation of “or”, which licenses the inference
that, according to the speaker, only one of the disjuncts is true, is due to a conversational
implicature. See §3.3 for further discussion.
16
Chapter 1
such an implicature. It is perhaps excusable to say that an implicature is
associated with a sentence (as uttered on a given occasion), as long as it is
firmly kept in mind that this is always to be thought of as informal shorthand:
it is the speech act, not the sentence, that carries the implicature. However,
it is wholly inexcusable to say that an embedded sentence, or any other expression smaller than a full sentence, may have conversational implicatures
associated with it. For example, the embedded sentences in (13a,b) (marked
by square brackets) could never give rise to conversational implicatures:
(13) a. If [C lives somewhere in the South of France], she must be bored.
b. There was a rumour that [Miss X produced a series of sounds that
corresponded closely with the score of “Home sweet home”].
A speaker who utters (13a) doesn’t state that C lives somewhere in the South
of France, and therefore in this context it will not have the implicature it has
in Grice’s example (11). In fact, no matter how loosely we allow ourselves to
speak of the connection between sentences and implicatures, there is no sense
in which this occurrence of “C lives somewhere in the South of France” can be
said to have implicatures at all. The same, mutatis mutandis, for (13b).
If embedded sentences cannot have implicatures associated with them, then
a fortiori words can’t have them, either. This may seem obvious enough, but
in view of later developments it is worthwhile to belabour this point a bit.
(14) Harry is either in Antwerp or in Brussels.
Typically, an utterance of (14) would convey that the speaker neither knows
whether Harry is in Antwerp nor whether he is in Brussels, and it has been
argued that these inferences are encoded in the lexical content of “or” (Zimmermann 2000, Geurts 2005). However, the standard view is Grice’s, who explains
these inferences as quantity implicatures, along the following lines. If the
speaker knew that Harry was in Antwerp, then he could have made a more
informative claim by saying, “Harry is in Antwerp”, but since he chose to make
a less informative and more prolix statement instead, he probably does not
know whether Harry is in Antwerp. By the same line of reasoning, the hearer
can infer that the speaker doesn’t know whether Harry is in Brussels.
Although I believe that Grice’s pragmatic analysis of “or” is right, and
I have my doubts about the lexicalist alternative, my main point, for now,
is just that there is no middle ground between these two views, given the
elementary fact that conversational implicatures are not carried by words.
To claim otherwise (as has been done by, e.g., Levinson 2000 and Chierchia
1.4 Generalised vs. particularised
17
2004) is quite simply an abuse of terminology. True, we have to allow for the
possibility that, as a language evolves, conversational implicatures turn into
conventional elements of content, and it might be argued that, in English, what
started out as a conversational implicature has become part of the lexical
content of “or”. But if this is true (which I doubt), then this inference has
ceased to be a conversational implicature.
Conversational implicatures are licensed not by words or sentences but by
actions of a certain kind, i.e. speech acts. This already suffices to ensure that
the etiology of conversational implicatures is far more general than that of, say,
lexical entailments or conventional implicatures. It is a fact about English “or”
that any English sentence of the form “ϕ or ψ” entails that at least one of ϕ and
ψ is true. By contrast, that utterances of disjunctive sentences may give rise to
ignorance inferences (“the speaker doesn’t know whether ϕ and doesn’t know
whether ψ”) is a fact about declarative statements with disjunctive content in
general, be they expressed in English, Basque, or Mandarin Chinese. Indeed,
Grice goes further even than this, in that he sees his theory of discourse as
just a special instance of a general theory of cooperative action.
This view has a very important methodological corollary: it entails that
if an inference can be explained in terms of conversational implicature, then
ceteris paribus such an explanation is to be preferred. This is what Grice calls
“Modified Occam’s Razor”: “Senses are not to be multiplied beyond necessity.”
(Grice 1978/1989: 47) Suppose, for example, that we had to adjudicate between the two theories of “or” discussed above: the lexicalist account which
says that the ignorance inferences associated with “or” are part of its lexical content and the Gricean account according to which these inferences are
conversational implicatures. Suppose, furthermore, that these two accounts
described the relevant facts equally well. Then we would still have to go with
the Gricean account, because it derives from general principles what according
to its lexicalist competitor has to be stipulated.
1.4 Generalised vs. particularised
One of the best-known distinctions introduced by Grice is one he isn’t too
forthcoming about: it is the distinction between particularised and generalised
conversational implicatures:
I have so far considered only cases of what I might call “particularized conversational implicature”—that is to say, cases in which an implicature is carried
18
Chapter 1
by saying that p on a particular occasion in virtue of special features of the
context, cases in which there is no room for the the idea that an implicature
of this sort is normally carried by saying that p. But there are cases of generalized conversational implicature. Sometimes one can say that the use of a
certain form of words in an utterance would normally (in the absence of special circumstances) carry such-and-such implicature or type of implicature.
(Grice 1975/1989: 37; emphasis added)
Grice uses two cases of quantity implicature to illustrate the distinction. One
of his examples is:
(15) X is meeting a woman this evening.
Grice observes that an utterance of this form “would normally implicate that
the person to be met was someone other than X’s wife, mother, sister, or perhaps even close platonic friend” (ibid.), adding that even when they are generalised, conversational implicatures do not thereby become conventional. And
that is more or less what Grice has to say about the generalised/particularised
distinction, leaving ample room for exegesis.
In my view, when Grice calls an implicature “generalised”, it is merely to
indicate that it tends to arise in such a broad range of contexts that it will
not seem to depend on the details of the context too much. But this is just to
say that the factors such conversational implicatures depend upon happen to
be shared by many contexts. For example, it seems highly likely that in just
about any context in which (15) is felicitously uttered, the interlocutors will
be interested in the identity of the woman in question, regardless what other
goals (if any) their discourse may serve, and in any such context, an utterance
of (15) will be liable to license the inference that the woman is not X’s wife,
mother, etc.
If my reading of Grice is correct, then the distinction between generalised
and particularised implicatures is “theoretically inert”, as Neale (1992: 524,
n. 18) puts it, which is in accordance with the fact that Grice has so little to
say about it. The reason why Grice introduces this distinction, I believe, is
that it is of strategic importance to him. One of the recurrent themes in Grice’s
work is that, in thinking about language, we have a tendency to overload our
words with meaning, in the sense that, whenever we observe that a word is
generally used in such-and-such a way, we are apt to put it down to the meaning
of the word (cf. §1.1). It is this tendency that has led some authors to suppose
that it is due to the lexical content of “or” that “ϕ or ψ” usually implies that
the speaker doesn’t know whether ϕ and doesn’t know whether ψ. One of
1.5 Cancellability
19
Grice’s main concerns is to show that this kind of reasoning is precipitate, and
that such inferences are better treated as conversational implicatures. And to
acknowledge the fact that such inferences can be pretty much standard, he
dubs them “generalised conversational implicatures”.
If this exegesis is on the right track, it is futile to argue against the generalised/particularised distinction, as some authors have done (e.g., Hirschberg
1985, Sperber and Wilson 1995), and it is mistaken to believe that generalised
implicatures should be allocated an entire level of their own in a theory of
interpretation (Levinson 2000). In general, it seems that Grice’s notion of
generalised conversational implicature has become undeservedly embroiled in
discussions it wasn’t designed for; more about which in Chapter 5.
1.5 Cancellability
Towards the end of “Logic and conversation”, Grice discusses a number of
“features” that are characteristic of conversational implicatures. For instance,
he says that conversational implicatures have to be “calculable” and that at
least some of them will be “non-detachable”. “Calculability” just means that
the hearer must be able to figure out how an implicature can be derived; which
is such a basic requirement that it hardly needs mentioning. An implicature
is “non-detachable” if it is dependent only on the content of the speaker’s
utterance, not its form. Hence, if an utterance of a sentence ϕ gives rise to an
implicature ψ, then ψ is non-detachable from what is said by means of ϕ if
alternative ways of expressing ϕ’s meaning would have given rise to the same
implicature. For instance, if we assume, if only for the purpose of illustration,
that “some” is synonymous with “at least two”, then we should expect that
(16a) and (16b) should give rise to the same quantity implicatures:
(16) a. Some of the stewardesses were snoring.
b. At least two of the stewardesses were snoring.
And indeed, both sentences will tend to convey that, according to the speaker,
not all the stewardesses were drunk. It should be obvious that not all conversational implicatures will be non-detachable in this sense, if only because Manner
implicatures do depend on the form of the speaker’s utterance. And there are
more exceptions, for as we will see in Chapter 6, quantity implicatures, too,
are sometimes contingent on linguistic form. Therefore, the diagnostic value
of non-detachability is rather limited.
20
Chapter 1
Another feature discussed by Grice is cancellability, and since this notion
figures rather prominently in the literature and is much more problematic
than it is usually given credit for, I will discuss it at greater length. According
to Grice, there are two ways of cancelling a conversational implicature. If a
speaker makes an utterance that gives rise (or might give rise) to an implicature
ϕ, he may add a rider like “but not ϕ”, in which case the cancellation is explicit,
or ϕ may be cancelled implicitly, by the context. (17a,b) are instances of
explicit cancellation:
(17) a. There is a garage round the corner, but it’s closed.
b. X is meeting a woman this evening—his sister, in fact.
Good examples of implicit cancellation are harder to find, but we encountered
one in the last section when discussing (18):
(18) Harry is either in Antwerp or in Brussels.
When uttered during an undoubtedly boring game in which the addressee
has to guess where Harry is, (18) will not prompt the ignorance inferences
otherwise associated with it. That is, it will not be construed as implying
that the speaker doesn’t know whether Harry is in Antwerp and doesn’t know
whether he is in Brussels, because the game presupposes that the speaker
knows where Harry is.
Cancellability is a useful diagnostic, because it can be used to separate
conversational implicatures from entailments and conventional implicatures.
To illustrate, compare (17a), which is felicitous, with (19), which is not:
(19) *There is a garage round the corner, but it’s not round the corner.
Since the propositional content of the first half of (19) is that there is a garage
round the corner, this proposition can only be cancelled at the price of selfcontradiction; which is why (19) is bad. Similarly, as we saw in §1.1, conventional implicatures cannot be cancelled with impunity, though the effect is not
as strong:
(20) ?Harry is rich but dull, though I wouldn’t want to suggest that there is a
contrast between these two properties. (= (2b))
If an inference is cancellable, it is neither a conventional implicature nor entailed by the sentence’s propositional content, and therefore it might be a
conversational implicature. However, it doesn’t have to be, since other types
of inference are cancellable, too:
(21) a.
Harry doesn’t realise October 2 was a Friday.
1.5 Cancellability
21
b. Harry doesn’t realise October 2 was a Friday, because it wasn’t: it
was a Saturday.
Ceteris paribus, an utterance of (21a) will suggest that Harry got promoted,
and it is widely agreed that, rather than a conversational implicature, this is
a presupposition triggered by the factive verb “realise”; (21b) shows that this
presupposition is cancellable. Hence, if an inference is cancellable, it may but
need not be a conversational implicature.
Even though cancellability may an effective heuristic for ruling out the
possibility that a given inference is either a conventional implicature or an entailment, it doesn’t reveal the true nature of the inference in question. Hence,
as a diagnostic test it is useful but less powerful than one might have hoped
for. What is more worrisome is that the word “cancellability” has proved to
be highly misleading. Grice’s explanation of cancellability, which is echoed
throughout the literature, is that conversational implicatures are cancellable
because a speaker may decide, or be forced to decide, to opt out of the Cooperative Principle. On this view, a speaker who utters (17a) or (17b) is being
uncooperative, if only in a small way. There are reasons for doubting that
this is correct. An utterance of (17a) may not be of immediate help to the
addressee, in the scenario envisaged by Grice, but arguably the speaker is
still providing topical information, which may prove useful in the longer term.
Hence, it is not so clear that someone who utters (17a) is being downright uncooperative. The same holds for (17b) and it holds a fortiori for (18). Indeed,
in cases like (18), in which an implicature is preempted by the context, it is
particularly doubtful that an inference is cancelled at all.7
Consider again the examples in (17). If these were instances of cancellation, we would have to say that in these cases the speaker first makes an
utterance that gives rise to a implicature ϕ, and then goes on to cancel ϕ.
This sounds somewhat perverse if not downright incoherent: a speaker who
invites his audience to draw an inference only to cancel it immediately afterwards can hardly claim to be cooperative. Perhaps, then, it is better to
say that the purpose of the second clause of (17a) is to prevent the hearer
from deriving an implicature he might have derived otherwise; and the same,
7.
To complicate things a bit further, there is an aspect of cancellability that many readers of
Grice seem to have overlooked, which is that Grice considers it to be a feature of generalised
conversational implicatures (Grice 1975/1989: 39). If this restriction is observed, then presumably (17a) will not count as an instance of cancellation, unless it can be argued that
“There is a garage round the corner” standardly gives rise to the inference that the garage
is open; which doesn’t seem very plausible.
22
Chapter 1
mutatis mutandis, for (17b). I believe that this makes more sense than the
standard cancellation view, because conversational implicatures are uncertain
in a way entailments and conventional implicatures are not. For one thing,
conversational implicatures are generally dependent on relevance, which is an
inherently fuzzy notion, not only because it is a matter of degree, but also because there will rarely be a perfect consensus, between interlocutors, on what
is relevant and what is not. For another, conversational implicatures involve
reasoning about the speaker’s intentional state (his beliefs, desires, and so on),
much of which is bound to be tentative. In contrast, given that it is part of
the lexical content of “but” that “ϕ but ψ” implies a contrast between ϕ and
ψ, there can hardly be any doubt that this is what the speaker means when
he utters a sentence of this form.
What I would like to suggest, therefore, is that within the general framework outlined by Grice, the notion that conversational implicatures are literally
cancellable is either misleading or misplaced. It is misplaced in cases like (18),
where a conversational implicature doesn’t arise, to begin with. It is misleading in cases like (17a) and (17b), which for reasons given in the last paragraph I
would prefer to analyse in terms of prevention rather than cancellation. Needless to say, none of this is to suggest that we should give up the cancellability
diagnostic, even if its value is somewhat limited, as I have argued. It is just
that the name of the diagnostic shouldn’t be taken too literally.
1.6 Gricean reasoning and the pragmatics of what is said8
Implicatures are suppositions made by the discourse participants in order to
make sense of each other’s speech acts. A speech act is a cooperative act in
much the same way as giving someone a light, assisting at an operation, or
playing the part of Fortinbras are cooperative acts, but speech acts are special
in that they carry propositional content: “what is said”, as Grice calls it. How
does this sort of content come about? Grice’s suggestion (and it is really no
more than that) is that, by and large, it is determined by the rules of the
language. Per linguistic convention, words and grammatical constructions are
assigned meanings, and the meaning of a complex expression is determined by
the meanings of its parts and the grammatical constructions they enter into.
Thus, the proposition expressed by “Snow is white” is built up from the lexical
8.
The second half of this section title was borrowed from Recanati (1989). See also Recanati
(2004).
1.6 Gricean reasoning and the pragmatics of what is said
23
meanings of the words “snow”, “is”, and “white”, pieced together according to
the conventions of English grammar.
If this was the whole story, we could say that what is said is conventional
simpliciter, but as Grice notes there is more to it.
(22) The ship left the lock.
Before this sentence can be assigned a truth value, we first have to decide: (i )
what is the time at which the purported event took place and (ii ) what are
referents of “the ship” and “the lock”; and in order to make that last decision,
(iii ) the word “lock” has to be disambiguated. That is to say, in order to
determine what is said, the hearer has to resolve any ambiguities the sentence
might contain, interpret tense, and assign referents to pronouns and definite
descriptions, as well as deal with a host of other aspects of interpretation
that Grice doesn’t mention. Clearly, such tasks may require that the hearer
take contextual information into account: what is said is not determined by
linguistic conventions alone, and contrary to what Grice seems to suppose, it
is well possible that what is said is radically underdetermined by linguistic
conventions.
In order to resolve ambiguities, referential terms, tense, and so on, the
hearer will have to ask himself what the speaker’s communicative intentions
are, and inevitably the Cooperative Principle will be involved in answering
this question. For example, the intended meaning of “lock” in (22) is probably
the one that is the most relevant, at the point in the discourse where this
sentence is uttered, so it would seem that the hearer needs to appeal to the
maxim of Relation in order to justify her choice. It appears, therefore, that
disambiguation has to rely on conversational implicatures. But how can that
be? Given that conversational implicatures are based on what is said, how
could they help to determine what is said? The Gricean project threatens to
become circular, or so it seems; Levinson (2000) calls this “Grice’s circle”.
Fortunately, however, it is easy enough to see that no such circle exists.
Consider example (22) again. How can the hearer determine the intended
meaning of “lock” before she has determined what the sentence means? Easy:
by asking herself what the sentence would mean if she chose one lexical meaning
rather than another. If “lock” referred to a mechanism for locking a door, the
propositional content of the sentence wouldn’t make much sense, whereas if
the word referred to a section of a canal in which the water level can be raised
and lowered, what is said would make perfect sense, and therefore the latter
meaning is preferred. The same, mutatis mutandis, for reference resolution,
24
Chapter 1
the interpretation of tense, and so on. Hence, “Grice’s circle” evaporates once
it is realised that sub-propositional interpretation issues can be resolved by
Gricean reasoning which is based not on what is said but on what would be
said if this or that interpretive choice were made.
Though “Grice’s circle” is a mirage, it brings us to a topic that deserves
some discussion. It so happens that Grice concentrates his attention on pragmatic inferences associated with utterances of single sentences.9 Due to this
restriction, which is by no means intrinsic to his programme, Grice’s claims
about conversational implicatures do not always generalise to all kinds of inferences driven by the Cooperation Principle. Put otherwise, Grice’s analysis
sometimes fails to carry over to all instances of what I will call “Gricean reasoning”. On the one hand, as we have already seen, some Gricean reasoning
is sub-propositional, which is to say say that it is concerned with the disambiguation of words, reference resolution, and so on. On the other hand, some
Gricean reasoning is supra-propositional: it is based on multi-sentence discourse segments. In the remainder of this section, I will say a bit more about
each of these levels.
Not all linguistic acts are sentence-sized. If we reserve the term “speech
acts” for those linguistic acts that are performed by means of sentences,10 there
are also linguistic acts which are expressed in units that are larger and others
that are expressed in smaller units. Beginning with the former, it is natural
enough to say that telling a story, giving a proof, pleading a case, exchanging
gossip, etc., are linguistic acts made up of several speech acts; and such largesized linguistic acts give rise to implicatures in much the same way as simple
speech acts do. To illustrate, consider the following recipe for negotiating
stairs with crutches, as provided by the American College of Foot and Ankle
Surgeons (www.footphysicians.com):
1. Seat yourself on a low step.
2. Move your crutches upstairs by one of these methods:
– If distance and reach allow, place the crutches at the top of the staircase.
– If this isn’t possible, place crutches as far up the stairs as you can—then
move them to the top as you progress up the stairs.
9.
As a matter of historical interest, it may be noted, though, that the first passage in Grice’s
work hinting at the Cooperative Principle concerns the resolution of ambiguities (Grice
1957). Cf. §8.5.
10. Though this usage is widely spread, some authors apply the term “speech act” more liberally.
A prominent case in point is Searle (1969), who treats referring as a speech act. I don’t have
any principled quarrels with this, but I do believe that such cases are sufficiently special to
deserve an appellation of their own.
1.6 Gricean reasoning and the pragmatics of what is said
25
3. In the seated position, reach behind you with both arms.
4. Use your arms and weight-bearing foot/leg to lift yourself up one step.
5. Repeat this process one step at a time. (Remember to move the crutches
to the top of the staircase if you haven’t already done so.)
Surely, this recipe would normally license the inference that the five steps have
to be executed in the order given (Manner), that there isn’t an extra step that
isn’t mentioned here (Quantity), and that the procedure will work (Quality).
It is clear, then, that extended stretches of discourse give rise to conversational
implicatures just like individual utterances do. There is nothing remarkable
about this, since the distinction between sentences and larger discourse units
is not very strict, to begin with. For example, stylistic considerations apart,
there doesn’t seem to be much of a difference between the sentences in (23a)
and (24a) and the mini-discourses in (23b) and (24b):
(23) a. Harry is dull but rich.
b. Harry is admittedly dull. However, he is rich, too.
(24) a. Tobago is primarily hilly and of volcanic origin.
b. Tobago is primarily hilly. It is of volcanic origin.
Just as there are linguistic acts beyond the sentence level, there are linguistic acts below that level. In using a definite description like “the ship”, the
speaker performs a referential act targeted at a particular individual; in using
a word like “lock”, he intends to convey a particular meaning; and so on. It
would be wholly remarkable if the hearer’s interpretation of such acts did not
proceed on the assumption that the speaker is trying to be cooperative. This
is not to imply that at the sub-propositional level the maxims are operative
in precisely the same way as they are on higher levels; since the maxims were
tailor-made for (propositional) speech acts, this is hardly to be expected. But
it is to say that we should expect to find implicature analogues below the
propositional level, and we do. For example, the use of a definite description
like “the ship” will normally be expected to be non-vacous (∼Quality) and
to refer to a relevant object (∼Relation). Furthermore, in general referential
expressions are expected to neither too general (∼Quantity 1) nor too specific
(∼Quantity 2), since lack of specificity will render it impossible to identify the
intended referent, while excessive specificity will result in moronic discourses
like the following:
(25) Once Little Red Riding Hood’s grandmother gave Little Red Riding
Hood a little cap made of red velvet. Because it suited Little Red Riding
26
Chapter 1
Hood so well, and Little Red Riding Hood wanted to wear it all the time,
Little Red Riding Hood came to be known as Little Red Riding Hood.
It is important to observe that there is a quite fundamental distinction
between the propositional and supra-propositional levels, on the one hand,
and the sub-propositional level, on the other, which lies in the fact that the
linguistic acts that make up the former are quite different from the ones that
define the latter. At the propositional level, the basic unit is the speech act, and
it is natural enough to extend this to the supra-propositional level: discourses
and discourse segments (paragraphs) are plausibly viewed as means for making
compound speech acts. However, linguistic acts at the sub-propositional level
are something different altogether. Referring to a ship and making a statement
about a ship are patently different things, and we would be ill-advised to ignore
the difference. If a referential expression α is insufficiently specific for the
hearer to determine the intended referent, we might say that the speaker’s use
of α violates an analogue of the first Quantity maxim, but not the maxim itself,
which was custom-made for speech acts. For these reasons, whereas I wouldn’t
mind extending Grice’s notion of implicature to the supra-propositional level,
I believe it would be a bad idea to apply this notion at the sub-propositional
level, as well.
To sum up, Gricean reasoning is employed not only to draw inferences
from what the speaker has said, but also to determine what he wants to say,
and contrary to what Levinson and others have claimed, this is not particularly problematic, though it usefully serves to highlight the fact that Gricean
reasoning may apply at various levels, which it is important to keep distinct.
In the following chapters, our attention will be focused on the more familiar
use of Gricean reasoning, i.e. on conversational implicatures proper, but the
sub-propositional level will loom large at the end of the book (§8.5).
The Standard Recipe
for Q-implicatures
2
Having surveyed the general framework of Gricean pragmatics, this chapter
homes in on the variety of conversational implicatures we will be concerned
with in the remainder of the book: quantity implicatures, or “Q-implicatures”
for short (Horn 1984). Our starting point is what I call the “Standard Recipe”
for deriving Q-implicatures. There are two reasons for calling it thus. First,
the literature has by and large confined its attention to inferences that follow
its pattern. Secondly, I suspect that in practice the Standard Recipe accounts
for many if not most Q-implicatures associated with declarative sentences.
However, standard though it may be, it doesn’t cover everything that might
justifiably be called a Q-implicature, even for declaratives.
The bulk of this chapter is devoted to fleshing out and amending the Standard Recipe, and this process will continue in the following chapters, where
the amendments become more substantial. In fact, I will end up arguing that
when push comes to shove the Standard Recipe gets things backwards: it
has all the right ingredients, but fails to combine them in just the right way
(Chapter 6). For the time being, however, we will be walking well-trodden
paths.
2.1 The Standard Recipe
I will introduce the Standard Recipe by way of an example. Clyde says:
(1)
Bonnie stole some of the pears.
On the classical Gricean account, (1) says that Bonnie stole at least some of
the pears, and may implicate that she didn’t steal all of them. This implicature
28
Chapter 2
is explained by assuming that the hearer reasons, and is entitled to reason, as
follows:1
i. Rather than saying (1), Clyde could have made a stronger statement:
(1*) Bonnie stole all the pears.
Why didn’t he do so?
ii. The most likely explanation is that Clyde doesn’t believe that (1*) is true:
¬BelC (1*).
iii. Clyde is likely to have an opinion as to whether (1*) is true: BelC (1*) ∨
BelC (¬(1*)).
iv. Between them, (ii ) and (iii ) entail BelC (¬(1*)): Clyde believes that Bonnie didn’t steal all the pears.
Note, to begin with, that this train of reasoning is conducted in terms of
the speaker’s beliefs. In the more recent literature on implicature, there is a
tacit consensus that this is right, but in the older literature, the same style
of analysis was often cast in terms of speaker’s knowledge (Gazdar 1979 and
Soames 1982 are prominent examples). I don’t see what could justify the
latter choice, and in fact I believe it’s plainly wrong, so I’m siding with the
majority view. It should be noted, however, that for our purposes nothing
hinges on the distinction between knowledge and belief (e.g., the logical form
of the reasoning displayed above is the same in either case), and since it is
sometimes more convenient to use the verb “know” rather than the weaker
“believe”, that’s what I will sometimes do.
The Standard Recipe starts with the observation that the speaker could
have made a stronger utterance than he did. Of course, since for any actual utterance there are always indefinitely many stronger utterances that in
principle could have been made instead, there will have to be substantial constraints on the alternatives available to the speaker at any given point in the
discourse. For example, if the speaker says ϕ and ψ is stronger than ϕ but also
considerably longer, then one good reason why the speaker didn’t bother to
say ψ is that the extra effort required wouldn’t have been worth the trouble.
Another, and much more important, constraint is that the alternatives to be
considered should be relevant to the interlocutors. Thus, in the case of (1),
“Bonnie stole some of the pears and today is Saint Genebald’s day”, though
1.
Readers who are unfamiliar with some of the notation appearing below are referred to the
overview of notation and abbreviations used in this book (p. 192).
2.1 The Standard Recipe
29
more informative than (1), would not count as an alternative.
Terminological note: Unless clearly indicated otherwise, I will use “alternative” as a theoretical term. An alternative is a sentence the speaker might
have uttered instead of the one he did utter. The exact import of “alternative”
will vary a bit, but its core meaning will remain constant: “the alternatives”
associated with a sentence, as uttered on a given occasion, will always be a set
of sentences.
Implicatures weak and strong Actually, the Standard Recipe involves not one
but two inferences that we might want to call “Q-implicatures”. The first one,
¬BelC (1*), is the weaker of the two, and it is important to realise precisely
what it says:
It is not the case that Clyde believes that Bonnie stole all the pears.
This does not entail that Clyde believes that Bonnie didn’t steal all the pears,
which is what the strong implicature says. ¬BelC (1*) allows for the possibility
that, as far as Clyde knows, Bonnie may or may not have stolen all the pears.
According to the stronger implicature, i.e. BelC (¬(1*)), Clyde has made up
his mind on this issue.
There is no established nomenclature for these two inferences. According to Soames, the weak one is a generalised implicature, while the second is
particularised, but as I have my doubts about the generalised/particularised
distinction (§1.4), I will not follow him in this point. Other options are to say
that only the first inference is a genuine implicature (the second inference derives from the first plus an auxiliary assumption), or to use Sauerland’s (2004)
distinction between “primary” and “secondary” implicatures. Since this is just
a terminological matter, and I happen to like the weak/strong distinction, I
will stick with that.
The strong implicature follows from the weak one plus the assumption that
Clyde either believes that Bonnie stole all the pears or believes that she didn’t.2
This assumption, too, has been given a variety of names, including “Experthood Assumption” (Sauerland 2004), “Authority Assumption” (Zimmermann
2000, Geurts 2005), and “Competence Assumption” (van Rooij and Schulz
2004). I will settle for the last one, but whatever its name, the underlying
idea is the same in any case; it is that, in general, implicature strengthening is
2.
As pointed out to me by Larry Horn, this style of disjunctive reasoning was already used
by Bartsch (1973) to explain “neg-raising” in attitude reports. See Horn (1989) for further
discussion.
30
Chapter 2
licensed by the assumption that the speaker knows what he is talking about.
In our example, this is to say that Clyde knows whether or not Bonnie stole all
the pears. Note, however, that this is slightly stronger than what is required
for obtaining a strong implicature: for that, we merely need to assume that
Clyde believes either that Bonnie stole all the pears or that she didn’t. That
is to say, the Competence Assumption only has to rule out the possibility that
Clyde hasn’t made up his mind as to whether (1*) is true: it has to eliminate
the dithering space between Clyde’s believing that Bonnie stole all the pears
and his believing that she didn’t. If Clyde was dithering on the matter, (1*)
and ¬(1*) would both be possible for him, and (2) would be true:
(2)
PossC (1*) ∧ PossC (¬(1*)), or equivalently: ¬BelC (¬(1*)) ∧ ¬BelC (1*)
The Competence Assumption is equivalent to the negation of (2). Hence, our
official definition of the Competence Assumption is somewhat weaker than
the intuitive notion. In general, this mild discrepancy will not matter; we will
briefly return to it in §2.3.
In a context that doesn’t support the Competence Assumption, the strong
implicature cannot be derived, and the Standard Recipe gives us only the weak
one. To illustrate how that might happen, consider (1) in two slightly different
scenarios. In both cases Clyde has witnessed Bonnie stealing some pears, but
while in scenario S1 he saw the platter from which she took the pears, in S2 his
view of the platter was obstructed by (say) a fat cat. Consequently, in S1 but
not in S2 Clyde knows whether any pears are left; or in other words, while the
Competence Assumption holds in S1 , it doesn’t hold in S2 , and the Standard
Recipe forecasts a strong implicature in the former case but not in the latter.
This seems to be correct: while in S1 we would infer from Clyde’s uttering (1)
that BelC (¬(1*)), S2 only allows for the weaker inference that ¬BelC (1*).
In my opinion, one of the main virtues of the Standard Recipe is that
it distinguishes between weak and strong implicatures, and connects them
via the Competence Assumption. Generally speaking, non-Gricean theories
are designed to account only for strong implicatures, which means that weak
implicatures are relegated elsewhere and the link between weak and strong
implicatures is severed (e.g., Gazdar 1979, Levinson 2000, Chierchia 2004,
2006). This is a less attractive solution than what the Gricean approach has
to offer. We will take up this issue in §7.3; the weak/strong distinction and
the Competence Assumption will be discussed at greater length in §2.3.
2.1 The Standard Recipe
31
The Standard Recipe Example (1) is a somewhat special case, in the sense
that its Q-implicatures are closely associated with one particular word, viz.
the scalar expression “some”; which is why it is called a “scalar implicature”. In
cases like (1), we can say that Q-implicatures arise because the speaker uses a
less informative word than he could have, and consequently his utterance could
have been more informative than it is. However, as Hirschberg (1985) was
the first to emphasise, Q-implicatures aren’t always like this, as the following
discourse illustrates:
(3)
A: Has Clyde’s book come out yet?
B: He has corrected the proofs.
B’s answer clearly implies that the answer to A’s question is no. This, too, is
a Q-implicature, which is obtained by applying the Standard Recipe:
i. Rather than giving the answer in (3), B could have made a stronger statement:
(3*) Yes, it has.
Why didn’t she do so?
ii. The most likely explanation is that B doesn’t believe that (3*) is true:
¬BelB (3*).
iii. B is likely to have an opinion as to whether (3*) is true: BelB (3*) ∨
BelB (¬(3*)).
iv. Between them, (ii ) and (iii ) entail BelB (¬(3*)): B believes that Clyde’s
book hasn’t come out yet.
The only difference between this derivation and the previous one is that the
relevant alternative, (3*), cannot be generated from B’s statement by substituting one word for another. It is important to see how small the difference is,
because it is often overrated, if only implicitly: since a very large part of the
literature on Q-implicatures confines its attention to the scalar variety, other
varieties have received less attention than they deserve, and there seems to be
a widespread tacit view that scalar implicature is so special that it is entitled
to a treatment of its own, which it has duly received at the hands of a long
list of distinguished authors, starting with Horn (1972). I consider this an
unfortunate development, and take it to be obvious that we should be aiming
for a more general account of Q-implicatures, in which scalar inferences come
out as just a garden variety. We will revisit this issue in the next chapter.
32
Chapter 2
Abstracting away from the peculiarities of the examples discussed in the
foregoing, here is the general form of the Standard Recipe for deriving Q-implicatures:
i.
ii.
iii.
iv.
S has said ϕ.
S could have made a stronger claim by saying ψ. Why didn’t he do so?
Presumably, it’s because S doesn’t believe that ψ is true: ¬BelS (ψ).
S has an opinion as to whether ψ is true: BelS (ψ) ∨ BelS (¬ψ).
Between them, (ii ) and (iii ) entail BelS (¬ψ): S believes that ψ is false.
A condensed form of the Standard Recipe was already given by John Stuart
Mill in the 19th century (and quoted on p. 2). More developed versions can be
found in Grice’s own work (e.g., Grice 1975, 1978) and the argument became
fully explicit at the hands of Soames (1982) and Horn (1989), and throughout
the more recent Gricean literature. Its formal properties have been studied by
van Rooij and Schulz (2004) and Spector (2006).
Non-declaratives One rather obvious limitation of the Standard Recipe is
that, like the maxims proposed by Grice, it is tailored to assertions. It is
not designed to work for questions, requests, commands, and so on. And despite the fact that the literature ignores them so completely that one might
easily come away with the impression that non-declarative speech acts don’t
have Q-implicatures, it’s not true that they don’t. In the previous chapter,
we saw that requests can have Q-implicatures (§1.3, example (12)). Here is
an example involving a question. Suppose that Clyde is collecting the ingredients for a cake he wants to bake. Bonnie enters the scene, and the following
exchange ensues:
(4)
B: Do you have all the ingredients?
C: Where’s the butter?
Clyde’s question may plausibly be taken to implicate that he did manage to
find the other ingredients, and this is arguably a Q-implicature, derivable as
follows:
i. Clyde could have asked, for example: “Where is the butter and where are
the eggs?” Why didn’t he do so?
ii. One possible explanation is that Clyde doesn’t need to be told where the
eggs are.
iii. And the most likely explanation for (ii ) is that he found them.
2.2 Inference to the best explanation
33
(And the same, mutatis mutandis, for the sugar, vanilla, and so on.) This is
a bit sketchy, admittedly, and fleshing out this story is not going to be trivial. But I take it that something along these lines will be needed to explain
the inference Clyde’s question gives rise to, and that the skeletal form of this
story is very similar to that of the Standard Recipe. In particular, both start
out with the question why the speaker failed to produce a more specific utterance, though the way this question is answered is not exactly the same for
interrogative and assertive speech acts.
2.2 Inference to the best explanation
Inferences and readings Implicatures are pragmatic inferences. Therefore, unless we are exceptionally liberal about what is to count as a “reading”, they
don’t affect the reading of a sentence. To illustrate this point, consider the
following sentences, all of which admit of several readings, and each in its own
way:
(5)
a. Clyde has a dog.
b. All the pears weren’t ripe.
c. Bonnie believes that she is pretty.
(5a) is ambiguous between a reading according to which Clyde owns a male
dog and one on which the animal in question may be any canis familiaris,
male or female. Similarly, (5b) has two readings: one which allows for the
possibility that some of the pears are ripe, and one which doesn’t. Finally,
(5c) is referentially ambiguous: the feminine pronoun may refer to Bonnie
herself or to another woman. In all these cases, the hearer is expected to select
the intended reading, using contextual cues as appropriate.
According to some authors, this is how (some) quantity inferences work, as
well. For example, Fox (2007) and Chierchia et al. (to appear) maintain that
(6a) is syntactically ambiguous between a strong reading that rules out (6b),
and a weak reading that doesn’t:
(6)
a. Bonnie stole some of the pears.
b. Bonnie stole all the pears.
The merits and demerits of this view (especially the latter) will be discussed
later (Chapter 7); the point I want to make now is just that it is fundamentally
different from the Gricean view. If the “not all” inference associated with
“some” in (6a) is an implicature, then the meaning of the sentence remains
34
Chapter 2
the same no matter whether the implicature is derived or not. The hearer
doesn’t select between a reading that includes the scalar inference and one
that doesn’t. Rather, he first selects a reading, and then decides whether or
not to derive the implicature.
Implicatures are abductive inferences Conversational implicatures are inferences of a special kind. To begin with, they are non-deductive. Clearly,
Clyde’s saying (6a) does not necessitate the conclusion that ¬BelC (6b). The
inference may be licit on many and perhaps even most occasions, but it isn’t
valid in the logicians’ sense of the word. What is it then? Implicatures are,
in Peirce’s (1931) terminology, abductive inferences, or inferences to the best
explanation, in Harman’s (1965) (I will use these terms interchangeably).3 If
Clyde utters (6a), and his interlocutors infer that ¬BelC (6b), then they do so
because they judge it is the best way of explaining why Clyde failed to make
a stronger claim.
A somewhat crude way of characterising abductive reasoning is that it uses
conditionals backwards: starting out with an explanandum ψ, you observe that
if any of ϕ1 , ..., ϕn were true, then ψ would follow, and then proceed to select
one among ϕ1 , ..., ϕn as the best explanation. This sort of diagnostic reasoning
is quite common in daily life. If I see that the milk bottle is empty, I’m likely
to conclude that somebody (a fellow member of my household, presumably)
drank it, although in principle it could have disappeared in countless other
ways: it could have been used for watering the plants, its contents could have
been transferred to another galaxy by aliens, and so on. Similarly, if a patient
exhibits all the symptoms associated with measles, his physician is likely to
infer that the best way of explaining the symptoms is by assuming that the
patient has measles (rather than, say, dengue fever). Incidentally, in medicine,
this mode of reasoning has been enshrined in the illustrious Zebra Principle:
“When you hear hoofbeats, think horses, not zebras.”
It is important to realise that abductive reasoning can go wrong without
therefore being wrong: it is a form of reasoning that tolerates exceptions. To
illustrate this characteristic feature, consider the following example. Having
observed the usual symptoms (a rough engine, steam escaping from the exhaust, etc.), a mechanic suspects that the Volkswagen he is looking at must
3.
See, e.g., Bach and Harnish (1979), Green (1996), Levinson (2000), and Chien (2008).
Thagard and Shelley (1997) criticise various attempts at defining abduction in terms of deduction. An excellent book-length treatment of inference to the best explanation is Lipton’s
(2004).
2.2 Inference to the best explanation
35
have a blown head gasket. But then when he looks under the bonnet, he discovers various pieces of equipment expressly designed to produce these symptoms;
there is nothing wrong with the engine. It turns out to be an elaborate prank.
Clearly, the mechanic’s diagnosis was incorrect: he thought the car had a blown
head gasket, and it hadn’t. Nevertheless, the man’s reasoning was impeccable:
this Volkswagen was merely an exception, not a counterexample.
The Gricean reasoning formalised in the Standard Recipe is very much
like that of the mechanic. With some ingenuity, we can concoct all sorts of
scenarios in which the Standard Recipe produces the wrong results. But they
are merely exceptions, not counterexamples. For instance, there are cases in
which Q-implicatures simply fail to arise. Consider (6a) once again, but now
in a context in which it has been discovered that the pears were poisoned,
and the only thing that matters is whether or not Bonnie took any of them.
In such a context, the truth value of (6b) is irrelevant, and it seems unlikely
that an implicature is intended by the speaker or inferred by the hearer (more
about relevance in §2.4).
Hence, in some cases, Q-implicatures don’t arise at all. More important
for our current purposes are cases where Q-implicatures do arise, but not in
the standard way. Consider the following example:
(7)
A: Clyde seems to have become such a penny-pincher. What’s the
matter with him?
B: He had to borrow money for some expensive pet he wanted to buy,
and now he’s saving wherever he can.
B’s use of the phrase “some expensive pet” may be construed as implicating
that she doesn’t know what species of animal Clyde bought. However, it may
also convey that she doesn’t care or even disapproves, perhaps because she
dislikes pets or Clyde’s infatuation with them. For reasons that aren’t quite
clear to me, being more general than might be expected may serve to express
a negative attitude, as in:
(8)
What’s this thing doing on the mantlepiece?
—where it is obvious to all concerned that the thing in question is a priceless
antique clock (cf. Cruse 1977). Perhaps the dismissive effect is caused by the
speaker’s studied indifference, but whatever the details of the explanation, it
is clear that these are Q-implicatures that don’t follow the Standard Recipe:
the question why the speaker chose to be uninformative is answered not by
assuming that there is something she doesn’t know but by assuming that
36
Chapter 2
she doesn’t care, or worse. Put otherwise, while in the Standard Recipe the
speaker’s key attitude is epistemic, in these cases it is evaluative.
Another type of case in which the Standard Recipe doesn’t apply is the
following. There is a popular stereotype according to which mothers are liable
to say things like:
(9)
I see that somebody hasn’t eaten his porridge yet.
In the kind of situation I’m alluding to, the mother is referring obliquely to
little Clyde (say), so she could have been more explicit by addressing him
directly:
(10) I see you haven’t eaten your porridge yet.
Still, although there is a stronger statement at hand, there is no implicature
to the effect that Clyde’s mother doesn’t believe that it is he who failed to
eat his porridge. I don’t know why mothers speak like this, but it is obvious
enough that it is a case of Q-implicature: the speaker deliberately avoids using
a (more informative) referential expression, and intends her audience to infer
whom she has in mind.
Another example in the same vein involves a stereotype of a more recent
vintage: it features two international terrorists calling each other on the phone.
One of them suspects that an “intelligence agency” may be tapping the line,
and says:
(11) A mutual acquaintance of ours will not be able to keep his appointment.
Like (9), the manifestly underinformative statement in (11) may serve to convey information about a particular individual, while at the same time the
speaker is communicating that it is unsafe to be more informative. The latter is clearly a Q-implicature, since it is the answer to the question why the
speaker chose to violate be less informative than he could have been; but the
Standard Recipe doesn’t apply here.
In examples like (9) and (11), it is clear that the Standard Recipe for
deriving Q-implicatures would produce the wrong results, for in both cases it
would be inappropriate to derive inferences of the form “The speaker doesn’t
know whether ...” Still, it is evident enough that these statements violate the
Quantity maxim, and thereby give rise to Q-implicatures. It is just that in
cases like these the best explanations are non-epistemic: they do not pertain
to what the speaker knows or doesn’t know.
Although some of the cases discussed in the foregoing are rather far-fetched,
that doesn’t affect my main point. Far-fetched or not, these examples tell us
2.3 Weak implicatures and competence
37
something important about Q-implicatures, and indeed about conversational
implicatures in general, namely that they heavily depend on world knowledge
in a way deductive inferences utterly fail to do. Though in many cases, the
Standard Recipe will give the best explanation for apparent violations of the
Quantity maxim, there are bound to be exceptions, too: cases in which other
explanations are better.
A Q-implicature is an answer to a particular type of why-question: “Why
didn’t the speaker say so-and-so?”, where “so-and-so” is a more informative, or
more specific, utterance than the one the speaker actually made. Nearly all
the work on Q-implicatures has confined its attention to one particular type
of answer to such questions, and I agree this is sensible enough: the cases
covered by the Standard Recipe are of central importance, and probably the
best place to start, so that’s what I will do, as well. Still, we shouldn’t lose
sight of the fact that there are other ways of deriving Q-implicatures. Rather
than an algorithm for producing Q-implicatures, the Standard Recipe is a
heuristic form of reasoning that is interesting and common enough to deserve
careful analysis. But there is only so much this analysis will be able to do. In
particular, we cannot hope to explain why, on any given occasion, the hearer
decides to adopt an explanation that follows the Standard Recipe, rather than
any of infinitely many alternative explanations that are always available in
principle.
2.3 Weak implicatures and competence
Competence and competence+ To the best of my knowledge, the first fully
developed version of the Standard Recipe was proposed by Soames (1982),
whose analysis differs from the one presented in §2.1 in two ways. One is
that, following Gazdar (1979), Soames casts his analysis in terms of knowledge
rather than belief; for reasons given before I will ignore this part of his proposal.
The second difference is that Soames uses a stronger version of the Competence
Assumption, which I will adorn with a plus to distinguish it from the one I
favour, and which (modulo the first difference) goes like this:
(12) Competence+ (Soames 1982)
If ψ then BelS (ψ), and if ¬ψ then BelS (¬ψ).
S is competent+ with respect to ψ if his beliefs agree with the facts; that is, if
S knows whether or not ψ is true. In Soames’s analysis, the Competence+
38
Chapter 2
Assumption replaces the Competence Assumption in the Standard Recipe
as presented in §2.1. Clearly, Soames’s notion of competence+ is stronger
than the unadorned one: assuming that ψ is either true or false, (12) entails that BelS (ψ) ∨ BelS (¬ψ), but not vice versa. Therefore, Soames-style
competence+ is stronger than it need be for strengthening weak implicatures.
By the same token, Soames’s analysis is less general. For consider Travis, who
has strong opinions on everything: give him any proposition ψ, and he will
either believe ψ or ¬ψ, regardless what the facts of the matter are. According to the Standard Recipe, Travis’s implicatures will always be strengthened,
which seems to be right; Soames doesn’t predict this.
Of course, Travis’s case is exceptional, if not pathological, and I wouldn’t
want to argue on this basis that the Competence+ Assumption is too strong.
In fact, I suspect that Soames is basically right in that generally speaking it
is competence+ , and not just competence, that is involved in the derivation of
strong implicatures. If I stick with the weaker version nonetheless, it is mainly
because it is simpler and more general: since it is defined in terms of knowledge
rather than (mere) belief, Soames’s notion imposes more stringent constraints
on the derivation of strong Q-implicatures than ours does, and although these
constraints may be satisfied in most cases, they are not needed in principle,
as the case of Travis illustrates.
What are weak implicatures? According to the Standard Recipe, Q-implicatures are derived in a number of steps, resulting initially in a weak implicature
of the form ¬BelS (ψ), which may be strengthened to BelS (¬ψ), if the Competence Assumption holds for ψ. For ease of reference, I give here a compressed
version of the argument:
i. Weak implicature: ¬BelS (ψ)
ii. Competence: BelS (ψ) ∨ BelS (¬ψ)
iii. Strong implicature: BelS (¬ψ)
The weak implicature is usually paraphrased as “The speaker doesn’t know
whether or not ψ is the case.” As before, I would prefer to talk in terms of
belief rather than knowledge: “The speaker has no opinion as to whether ψ
is true or false.” But the difference doesn’t really matter, because either way
the paraphrase is stronger than the inference licensed by the Standard Recipe.
For, the paraphrase is captured by the conjunction ¬BelS (ψ) ∧ ¬BelS (¬ψ),
which says that the speaker believes neither ψ nor ¬ψ, and therefore entails,
but isn’t entailed by, ¬BelS (ψ).
2.3 Weak implicatures and competence
39
Before we continue, it may be helpful to give the following diagram, which
displays how Competence and the various candidate implicatures relate to
each other:
non-belief
BelS (ψ)
z
BelS (¬ψ)
belief
disbelief
|
¬BelS (ψ)
}|
{
¬BelS (ψ) ∧ ¬BelS (¬ψ)
no opinion
{z
}
BelS (ψ) ∨ BelS (¬ψ)
competence
This is to be understood as saying that “non-belief” is the disjunction of “disbelief” and “no opinion”, while “competence” is the disjunction of “belief” and
“disbelief”. Note, furthermore, that “no opinion”, i.e. the intuitive paraphrase
of the weak implicature, is the negation of “competence”:
“¬BelS (ψ) ∧ ¬BelS (¬ψ)” is true iff “BelS (ψ) ∨ BelS (¬ψ)” is false.
Now, which of the two is the true weak implicature: ¬BelS (ψ), which is
what the Standard Recipe gives us, or the stronger ¬BelS (ψ) ∧ ¬BelS (¬ψ),
which corresponds to the standard way of paraphrasing the weak implicature?
According to Pouscoulous (2006), who was the first to broach this issue, it is
the latter. To explain why, suppose hearer H has executed the first step of
the Standard Recipe, deriving ¬BelS (ψ). Then, in principle, there are three
options for H to choose from:
i. H may decide not to adopt the Competence Assumption, and is then left
with just ¬BelS (ψ).
ii. H may decide that the Competence Assumption is true, and then BelS (¬ψ).
iii. H may decide that the Competence Assumption is false, and then ¬BelS (ψ)
∧ ¬BelS (¬ψ).
To make this more concrete, we return to our original example: Clyde says:
(13) Bonnie stole some of the pears.
Given that Clyde could have said,
(13*) Bonnie stole all the pears.
the three options distinguished above amount to the following:
40
Chapter 2
i. ¬BelC (13*): Clyde either believes (13*) is false or he believes (13*) may
or may not be true.
ii. BelC (¬(13*)): Clyde believes (13*) is false.
iii. ¬BelC (13*) ∧ ¬BelC (¬(13*)): Clyde believes that (13*) may or may not
be true.
Pouscoulous considers it highly unlikely that by saying (13) Clyde should have
wanted to communicate merely that (i ) is the case, and I believe she is right
about this: presumably, what Clyde intended to communicate to his audience
was either (ii ) or (iii ). However, taking the hearer’s point of view, it may be
doubtful what the status of the Competence Assumption is, so in at least some
contexts (i ) may be all the hearer is able to infer with reasonable confidence,
in which case he may want to ask for clarification:
(14) What do you mean: she didn’t steal all of them or you don’t know?
So, although hearers may prefer to either adopt or reject the Competence
Assumption, they may sometimes remain undecided, and be left with a rather
weak construal of what the speaker meant to convey.
The upshot of the foregoing discussion is as follows. Although the Standard
Recipe isn’t wrong, it is incomplete, and the common way of presenting it is
not entirely correct. Here is how it should go. Having derived an inference of
the form
(15) ¬BelS (ψ)
the hearer may go on to assume either that the Competence Assumption holds
or that it doesn’t hold. In the former case, it follows that
(16) BelS (¬ψ)
In the latter case, the inference is that
(17) ¬BelS (ψ) ∧ ¬BelS (¬ψ)
The problem with the standard way of presenting the Standard Recipe is that
(15) is offered as meaning (17), which it doesn’t, since the former is weaker
than the latter. Note that (16) and (17) are logically independent (neither
entails the other), which is to say that although (16) may seem to be stronger
than (17), it isn’t. Therefore, we can rule out the possibility that (16) is
derived by strengthening (17): it can only derive from (15).
Having made the requisite distinctions, I will in the following take the
liberty of not being too strict about them, the reason being simply that for
most of the discussion the difference between (15) and (17) is immaterial.
2.3 Weak implicatures and competence
41
Disjunction and ignorance Disjunctive sentences of the form “ϕ or ψ” generally give rise to the implicature that the speaker doesn’t know which of ϕ and
ψ is true. For example, if Clyde says (18), his utterance is likely to be construed as implying that he is agnostic about the truth values of its disjuncts,
(19a,b):
(18) Bonnie stole an apple or a pear.
(19) a. Bonnie stole an apple.
b. Bonnie stole a pear.
Grice’s (1978/1989: 46-47) explanation for these ignorance inferences is seemingly straightforward: had Clyde known that (19a) is true, he should have said
so, and since he didn’t he is implicating that ¬BelC (19a); the same for the
(19b). The reason why this is only seemingly straightforward is that, according
to the Standard Recipe, the next step should be to strengthen these implicatures, which would lead us to conclude that BelC (¬(19a)) and BelC (¬(19b)):
Clyde believes both disjuncts to be false. Assuming that Clyde is not violating the Quality maxim, he believes what he says, so we cannot infer both
BelC (¬(19a)) and BelC (¬(19b)); but why not infer one of the two? The
answer is that we cannot make the Competence Assumption for either disjunct, on pains of contradiction (Sauerland 2004). Still assuming that Clyde
is not violating the Quality maxim, he believes what he says, and therefore
BelC ((19a) ∨ (19b)); so, taken together with the weak implicatures we have
already, this yields:
i. BelC ((19a) ∨ (19b))
ii. ¬BelC (19a)
iii. ¬BelC (19b)
[Quality]
[Quantity]
[Quantity]
If, in addition, we adopt the Competence Assumption for (19a), we get a
contradiction, as follows:
iv.
v.
vi.
vii.
BelC (19a) ∨ BelC (¬(19a))
BelC (¬(19a))
BelC (19b)
contradiction
[Competence]
[from (ii ) and (iv )]
[from (i ) and (v )]
[between (iii ) and (vi )]
Of course, assuming competence for (19b) will yield the same result. In short:
provided Clyde believes what he says, the weak implicatures of his utterance
imply that the Competence Assumption fails for both disjuncts, and therefore
no strong implicatures can be inferred. We might say, therefore, that in the
case of disjunction, weak implicatures are non-competence inferences.
42
Chapter 2
Competence by default? It is sometimes suggested that hearers adopt the
Competence Assumption by default. Very roughly, the idea is that, in the absence of evidence to the contrary, hearers will assume that the speaker knows
what he is talking about. Thus formulated, the claim is plausible enough,
though it is less than self-evident what it means for Q-implicatures. I propose
to read it as follows:
(20) Competence by default
If the hearer applies the Standard Recipe, and asks himself why the
speaker uttered ϕ instead of the stronger ψ, he will assume by default
that the speaker is competent with respect to ψ.
If this is correct, we predict that, if the Standard Recipe is deployed, there will
be an inference of the form BelS (¬ψ) by default. Note that this is superficially
similar to, but really very different from, the claim that certain expressions
trigger default inferences of the form BelS (¬ψ), a claim that will be discussed
at length in Chapter 5.
Another way of formulating (20) is that, generally speaking, a speaker who
has sufficient evidence for making a claim ϕ will be taken to also have sufficient
evidence for deciding on the truth value of ψ, where ψ is a stronger alternative
to ϕ. This seems plausible to me, though of course it may turn out to be
false. The only way to settle this issue, I suspect, is by collecting quantitative
data, but unfortunately I don’t see how this might be done. In the meantime,
I suggest that we accept (20) as a working hypothesis.
2.4 Relevance
The derivation of a Q-implicature starts with the observation that the speaker
could have made a stronger statement than he did, and then proceeds to
explain why he didn’t do so. It is easy enough to see that it would be a bad
idea to raise this question for every stronger statement the speaker could have
delivered, not only because doing so would cause a dramatic slowdown of the
interpretation process, but also because it would produce absurd results like
the following. Bonnie says:
(21) Clyde is a dope.
Instead of saying (21), Bonnie could have a made the following statement,
which is patently stronger:
2.4 Relevance
43
(22) Clyde is a dope and Jupiter is larger than Saturn.
So, apparently Bonnie doesn’t believe that (22) is true, and since she has
committed herself to the truth of the first conjunct, we are forced to conclude
that she doesn’t believe that Jupiter is larger than Saturn.
Something must have gone wrong, and it is fairly evident what it is: not
any old statement that happens to be stronger than (21) qualifies as a relevant alternative. Indeed, for practically any given statement and context,
the overwhelming majority of stronger statements fail to be relevant. Thus
the question arises: how to separate the relevant alternatives from the irrelevant ones? This question is ... uhm ... difficult, and since I have little hope of
advancing the state of the art in this respect, the following remarks serve to
outline the problem more than anything else.
Discourse purposes Let’s start with Grice’s own formulation of the first Quantity maxim:
Make your contribution as informative as is required (for the current purposes
of the exchange). (Grice 1975/1989: 26)
This formulation suggests that according to Grice the only relevant alternatives
are those that would have contributed to the purposes of the discourse (though
elsewhere he seems to adopt a more liberal view on relevance; e.g., Grice
1975/1989: 38). This is an appealing idea, and to illustrate how it might
work, I will use one of Grice’s own examples:
(23) The prize is either in the garden or in the attic. (Grice 1978/1989: 45)
As explained in the last section, an utterance of (23) will tend to implicate that,
as Grice puts it, the speaker has “non-truth-functional grounds” for believing
that his statement is true; which is Grice’s way of saying that (23) will typically
license the implicature that the speaker doesn’t know which of the following
is true:
(24) a. The prize is in the garden.
b. The prize is in the attic.
But now suppose (23) is uttered in the context of a treasure hunt, by a parent
who knows full well where the prize is; then there will be no such implicature,
obviously. In such a case, neither (24a) nor (24b) count as stronger claims a
cooperative parent could have made, in view of the purpose of the exchange,
because either statement would have spoiled the game.
44
Chapter 2
What are discourse purposes? This, too, is a hard question, but part
of the answer must be that discourse purposes are common goals, i.e. goals
that are shared between speakers. All parties can come to the discourse with
any number of private goals, but only some of these will become discourse
goals, and not every goal the discourse helps to achieve is a discourse goal.
For example, even if my main reason for steering the conversation towards
minimalist syntax is that I want to impress my audience with my knowledge
of the topic (the example is fictitious), that would not be a common objective,
and therefore not a discourse goal. Rather, the purpose of the discourse would
be to exchange information about minimalist syntax.
A common way of setting up discourse goals is by asking questions. For
instance, by asking,
(25) Where’s the Earl Grey?
Bonnie is making a bid for establishing a discourse goal of a very particular
kind, viz. that she should learn where the Earl Grey is. If Clyde accepts this
as the next discourse purpose on the agenda, then he should undertake to
answer the question to the best of his ability; for instance, by offering one of
the following:
(26) a. In the blue cupboard.
b. On the top shelf of the blue cupboard.
Now (26b) is more informative than (26a), and therefore an utterance of the
former might well implicate that Clyde cannot be more precise. However, if it
is agreed that (26a) is good enough for achieving the current discourse goal,
then no Q-implicature will arise.
An illustration of how implicatures are constrained by questions is offered
by the following minimal pair:
(27) A: Has your article been published yet?
B: I corrected the proofs.
(28) A: What was your contribution to that article?
B: I corrected the proofs.
Whereas B’s statement in (27) may be interpreted as implicating that for all
B knows, the article hasn’t been published yet, the same statement in (28) is
unlikely to give rise to such an implicature. What explains the contrast is that
the discourse goals are different.
2.4 Relevance
45
A question serves to set up a discourse goal which its answer should help to
achieve, and thus helps to establish what is relevant in the following. Hence,
questions are employed for negotiating relevance, and this observation provides
us with a useful tool for studying the effects of relevance on implicatures, as well
as other aspects of interpretation: by observing how the interpretive effects of
a statement change depending on what questions it might be used to address
(as we did in the foregoing examples), we are in effect varying the relevance
factor.
Hearer’s interest If it was determined entirely by the immediate purposes
of the discourse, the notion of relevance would be relatively unproblematic.
Unfortunately, however, it isn’t.
(29) A: I tried to call you yesterday afternoon. Where were you?
B: I was playing tennis with a friend of mine.
B’s utterance is likely to implicate that his tennis partner, Clyde, is not among
A and B’s mutual acquaintances. But this information is not relevant to the
immediate purpose of the discourse, which is to answer A’s question, and thus
explain why B wasn’t answering the phone. In cases like (29), we probably
have to say that relevance is determined by the hearer’s interests. B will be
aware that Clyde’s identity is of potential interest to A if Clyde happens to
be a mutual friend, A’s employer, the Dalai Lama, etc., and this is enough
to make relevant some of the alternative statements that are stronger than
(29B).4
General interest We have seen that relevance is determined not only by the
immediate purposes of the discourse, but by the hearer’s interests, as well.
And that’s not the end of it, as the following dialogue illustrates:
4.
Although discourse purposes and hearer’s interests will typically overlap, they are not the
same. For example, even if it is common knowledge between a journalist and a politician
she is interviewing that neither has any personal interest in the other, whatsoever, this will
not prevent them from producing all manner of implicatures.
It is the hearer-centred notion of relevance that takes pride of place in Sperber and
Wilson’s work (e.g. Sperber and Wilson 1995, Wilson and Sperber 2004), though I should
note that Sperber and Wilson’s use of “relevance” doesn’t align with any intuitive notion of
relevance, as far as I can tell. Rather, what they call “relevance” is the ratio between hearer’s
relevance and processing effort; so the unit of measurement of relevance à la Sperber and
Wilson is common relevance per kilojoule.
46
Chapter 2
(30) A: Where is Bonnie?
C: She went out to buy a piece of furniture.
C’s answer is very likely to implicate that he doesn’t know what kind of furniture Bonnie went out to buy (or that he doesn’t care, cf. the discussion of
(7), §2.2). This information does not seem relevant to the immediate purpose
of the discourse, and though it may be relevant to A’s personal interests, it
doesn’t have to be. For suppose that it is mutual knowledge between A and C
that A has no interest in furniture whatsoever, and couldn’t care less whether
Bonnie bought a sofa, a dinner table, or a rocking chair.5 Then, still, the
question is bound to arise why C chose to be so unspecific. Note that if C had
said (31) instead,
(31) C: She went out to buy a sofa.
there wouldn’t have been an implicature to the effect that, e.g., B didn’t know
if B went out to buy a blue sofa. In this respect, there is a sharp contrast
between (30C) and (31C), as possible answers to A’s question. The same
contrast can be observed in the following pair:
(32) a. I saw an animal on the lawn this morning.
b. I saw a dog on the lawn this morning.
Whereas (32a) seems quite liable to implicate that the speaker isn’t sure what
species of animal he saw, it would require special circumstances for (32b) to
implicate that the speaker isn’t sure what breed of dog he saw (cf. Brown 1958,
Cruse 1977, Rosch 1978, Hirschberg 1985).
The examples in (30)-(32) indicate that there are general expectations to
the effect that, when introducing a new discourse entity, speakers should employ expressions of at least a minimum level of specificity: “sofa” and “dog” are
sufficiently specific; “piece of furniture” and “animal” are not. These expectations are general in that they don’t fluctuate wildly from context to context
5.
On the other hand:
Information, like money, is often given without the giver’s knowing to just what use
the recipient will want to put it. If someone to whom a transaction is mentioned
gives it further consideration, he is likely to find himself wanting the answers to
further questions that the speaker may not be able to identify in advance; if the
appropriate specification will be likely to enable the hearer to answer a considerable
variety of such questions for himself, then there is a presumption that the speaker
should include it in his remark; if not, then there is no such presumption. (Grice
1975/1989: 38)
2.4 Relevance
47
or from hearer to hearer, though they may be culture specific. This point was
first made by Matsumoto (1995), on the basis of evidence from Japanese and
Korean.
Matsumoto observes that, in English, a sentence like (33) will not tend to
implicate that the speaker doesn’t know whether Michio is older or younger
than Takashi:
(33) Michio is Takashi’s brother.
However, if (33) is translated into Japanese, the result is a sentence that does
have this implicature:
(34) Kochira wa Takashi-kun no kyoodai no Michio-kun desu.
this
top Takashi-Mr. gen brother gen Michio-Mr. cop
The contrast between the two languages is explained by the fact that, in
addition to words corresponding to English “brother” and “sister”, Japanese
also has a system of sibling terms that encode sex as well as seniority:
ani “older brother”
ane “older sister”
otooto “younger brother”
imooto “younger sister”
Moreover, and crucially, these are the preferred terms for introducing siblings,
and therefore the use of “kyoodai” in (34) is marked in much the same way
that “animal” is marked in (32a). Quite possibly, this preference is due to the
fact that age differences play a more important role in Japanese culture than
they do in other parts of world (see §6.5 for further discussion).
To recapitulate: If a speaker utters a sentence ϕ when a stronger sentence
ψ could have been used instead, the difference doesn’t matter (i.e., will not
trigger Q-implicatures) unless it is relevant, and it can be relevant in at least
three ways. The difference between ϕ and ψ may be (i ) relevant to the purposes
of the discourse, (ii ) of potential interest to the hearer, or (iii ) of general
interest. These three options need not be disjoint, but they are not necessarily
the same, either.
Relevance isn’t one thing: there are a number of disparate factors that
determine what is relevant and what is not, and while some of these factors
are context dependent to a high degree, others are less so. Unfortunately,
although relevance is of crucial importance to an adequate understanding of
conversational implicatures of all stripes, I will have very little to add to the
preliminary observations offered in this section. Relevance-related considerations will loom large at many points in the book, but I don’t have anything
48
Chapter 2
like a theory of relevance to offer; that is to say, a theory that is sufficiently
precise, not overly simplistic, and that draws together the foregoing observations with many others that have been made in the literature. I’m not aware
of any theories that meet these basic requirements, and, sadly, I doubt that I
will see one in my lifetime.6
2.5 Conclusion
In this chapter, we have studied the Standard Recipe, a form of reasoning that
gives rise to two types of Q-implicatures, weak and strong, which are connected
by the Competence Assumption. It is important to bear this general pattern
in mind, not only because it will recur again and again, but also because, as
we will see in §7.3, alternative theories that seek to explain these inferences in
non-pragmatic terms can only account for strong implicatures and have to let
the weak ones fall by the wayside.
More importantly, however, the Standard Recipe is a specimen of inference
to the best explanation, which is to say that it is not the only way of deriving Qimplicatures and, a fortiori, it is not an algorithm for deriving Q-implicatures.
The fact that Gricean reasoning is essentially abductive in nature is often
overlooked, by friend and foe alike, and all too often the Standard Recipe is
treated as if it were a logical calculus; which it isn’t. This trend is particularly
strong when the focus is on scalar implicatures, a sub-class of Q-implicatures
to which we now turn.
6.
This may seem like a slur on theories in which relevance plays a central role, but that’s
not how it is intended. It’s just that, as far as I can tell, all such theories presuppose some
notion of relevance, which is left unaccounted for.
Scalar implicatures
3
By far the most popular variety of Q-implicatures are scalar implicatures,
which were put on the pragmatic map by Horn (1972). Most of the examples
we discussed in the foregoing were of this sort; (1) is another one:
(1)
Wilma read some of the papers.
(2)
a. Wilma read many of the papers.
b. Wilma read most of the papers.
c. Wilma read all of the papers.
The distinctive feature of scalar implicatures is that we can use lexical substitution to generate the relevant alternatives from the sentence uttered. Thus,
stronger alternatives for (1) would be (2a-c), which are obtained simply by
replacing “some” with “many”, “most”, and “all”, respectively. The idea is that
there is a scale, say hsome, many, most, alli, which orders its elements from
weaker to stronger, and which in effect generates the alternatives to a given
sentence. Hence, using this substitution method, an alternative to a given
sentence ϕ is generated by replacing an expression α in ϕ by a stronger scalemate of α’s. This generative view, as I propose to call it, originated in Horn’s
(1972) seminal work, and was subsequently adopted by Gazdar (1979), Soames
(1982), Hirschberg (1985), and many others. More recently, the view is still being taken for granted, e.g., by Sauerland (2004), Chierchia (2004), and Chemla
(2007).
In this chapter, I will first expound and then criticise the generative approach; the latter on two main counts. First, I will argue that Horn scales are
redundant: we don’t need a scale to tell us that “all” is stronger than “some”,
for this much follows from the meanings of the words, and if we really want to
use the ordering encoded in the scale, we get into trouble. Secondly, I will try
to show, using the case of disjunction, that the class of scalar implicatures, as
50
Chapter 3
staked out by the substitution method, does not consitute a natural kind: the
generative view is compelled to draw boundaries that shouldn’t be there.
Before getting down to business, I should like to stress that, though his
name will occur rather frequently in the coming pages, this chapter is not
a critique of Horn’s views, past or present. In fact, I won’t be criticising
anybody in particular. The version of the substitution method that will serve
as my scapegoat is extremely simple, and is not supported by anyone anymore.
However, that doesn’t really matter for the main points I want to make.
3.1 Horn scales and the generative view
Let’s start by introducing a notational device that will make it easier to discuss
sentences and their alternatives: ϕ[...] is a sentence frame in which the dots
mark an open position, for example “Wilma read [...] of the papers.” If ϕ[...]
is a sentence frame, and α is an expression of the appropriate type, then ϕ[α]
is that sentence in which α fills the open position in ϕ[...]. Hence, ϕ[α] and
ϕ[β ] are the same except that the latter has β where the former has α.
A Horn scale is simply a sequence of increasingly informative expressions.
Examples are:
(3)
a.
b.
c.
d.
hsome, many, most, alli
hwarm, hot, scaldingi
hclever, brillianti
hor, andi
Two expressions will be called “scalemates” iff there is a Horn scale that contains them both; β is a stronger scalemate of α’s iff there is a Horn scale in
which β occurs to the right of α.
The elements of a Horn scale are ordered according to how informative
they are.1 For example, everyone who is “brilliant” is “clever”, as well, but not
vice versa; or in other words, the set of brilliant people is a proper subset of
the set of clever people.2
1.
I will be assuming throughout that informativeness is defined in terms of entailment. There
has been some discussion in the literature about so-called “non-entailment scales” (see especially Hirschberg 1985). For reasons that will transpire in due course, I believe this is a
non-issue.
2. Note that this is not an empirical claim about the meanings of words like “clever” and “brilliant”. More accurately, it is not a claim that can be verified directly, for to call somebody
“clever” often is to imply that the person in question is not brilliant. However, since we
3.1 Horn scales and the generative view
51
With these preliminaries out of the way, we proceed to define a very simple
procedure for generating the alternatives for a given sentence, using Horn
scales:
(4)
Using Horn scales to generate alternatives
ϕ[β ] is an alternative to ϕ[α] iff β is a stronger scalemate of α’s.
Assuming that hsome, many, most, alli is the only available scale, this yields
the following results for the sentences in (1) and (2):
(2a), (2b), and (2c) are the alternatives to (1).
(2b) and (2c) are the alternatives to (2a).
(2c) is the sole alternative to (2b).
There are no alternatives to (2c).
When the Standard Recipe is applied to the alternatives of sentence (1), for
example, this yields the following weak implicatures, each of which may be
strengthened, as usual, provided the Competence Assumption holds:
¬BelS (Wilma read many of the papers)
¬BelS (Wilma read most of the papers)
¬BelS (Wilma read all of the papers)
The exclusive interpretation of “or” is explained along the same lines:
(5)
a. Barney has bought a Manet or a Monet.
b. Barney has bought a Manet and a Monet.
On the classical, inclusive interpretation of “or”, the truth-conditional meaning
of (5a) merely rules out the possibility that Barney has bought neither a Manet
nor a Monet: the sentence is true if Barney has bought a Manet or a Monet or
both. Therefore, we can assume that “or” and “and” share a scale and that the
latter is stronger than the former, and then the substitution method predicts
that (5b) is an alternative to (5a). If we put this alternative through the
Standard Recipe, we get two grades of exclusivity implicature: ¬BelS (5b)
(weak) and BelS (¬(5b)) (strong). Thus the exclusive construal of (5a) is
explained as an ordinary scalar implicature.
We now have an apparatus that has been viewed as something I have argued
could not exist: a mechanical procedure for generating implicatures. On this
view, there is an algorithm for churning out scalars implicatures, after all, and
it consists of the following components:
want to treat this as a conversational implicature, it had better not be entailed by the lexical meaning of “clever”. Hence, to say that “brilliant” is stronger than “clever” is to adopt
an hypothesis that can only be tested as part of a larger theory of interpretation.
52
Chapter 3
i. a set of Horn scales
ii. a procedure for generating alternatives
iii. the Standard Recipe
It is clear why the substitution method has proved to be so popular: it is
intuitively appealing, easy to formalise, and simple. And it is even more
appealing to suppose that this method is in fact an algorithm for producing
a special class of Q-implicatures, viz. scalar ones. There are two reasons for
doubting that this could be right. One is that implicatures are inferences
to the best explanation, which means that even if a procedure like this were
correct in many and perhaps even the majority of cases, it is bound to fail in
some (cf. §2.2). Another is that implicatures are contingent on considerations
of relevance, which, prima facie at least, don’t seem to be taken into account
by the generative view (cf. §2.4). Hence, two perfectly sound a priori reasons
for rejecting this position out of hand.
However, although the abduction argument is incontrovertible, as far as I
can see, the relevance argument may not be. For, it might be argued that,
contrary to what I have alleged, the generative view already takes into account
relevance: it is encoded in Horn scales. To explain how, let us turn to a
question that has received remarkably little attention in the literature: What
are Horn scales, anyway? I believe I can answer that question, and I suspect
that my answer will satisfy most scale theorists (though I’m not sure). Briefly,
what I would like to suggest is that Horn scales represent hearers’ expectations
in the sense that if β is a scalemate of α’s, β is taken to be relevant in a context
in which α is used.
To explain this a bit further, let’s start by observing that, although Horn
scales are ordered in terms of semantic strength, most sequences of increasingly
informative expressions do not qualify as Horn scales. To see why, consider
the following candidate scales:
(6)
a. hanimal, dogi
b. hanimal, dog, schnauzeri
If (6a) is a bona fide Horn scale, we predict that (7) should give rise to the
implicature that the speaker doesn’t know (or, perhaps, didn’t know at the
time) whether or not the sighted animal was a dog:
(7)
I saw an animal on the lawn this morning.
This prediction is plausible enough. However, if (6b) was a Horn scale, as well,
we would make an analogous prediction for (8): it should yield the implicature
3.1 Horn scales and the generative view
53
that the speaker doesn’t (or didn’t) know whether or not the dog in question
was a schnauzer:
(8)
I saw a dog on the lawn this morning.
This prediction is not nearly as plausible, and if we want to avoid it, we will
have to assume that (6b) is not a Horn scale.
On the interpretation I’m suggesting, if (6b) is a Horn scale while (6a) is
not, then it is because hearers have different sorts of expectations about the
words (or concepts) “animal” and “dog”: if an individual x is introduced as an
“animal”, the hearer is likely to wonder what species of animal x is, whereas
if x is introduced as a “dog”, the question whether x is a schnauzer will not
in general present itself with equal force. Of course, such expectations may
be dependent on the context to a greater or lesser degree. For instance, if (8)
occurred in a conversation between avid dog lovers, it is quite possible that the
hearer would ask herself whether or not the dog in question was a schnauzer,
and this could be represented by assuming that, in the given context, (6b) is
a Horn scale.
If this is the right way of viewing Horn scales (and it’s the best I can think
of), such scales are partial representations of contextual relevance. Some of
these representations, like (6a), will be fairly constant across contexts; others,
like (6b), will be more dependent on specific contextual features. Either way,
it will be clear that Horn scales don’t explain what relevance is; they merely
provide a format for representing it. It will also be clear that this format will
not be suitable for representing relevance in general. For example, the way
relevance expectations are affected by questions (§2.4) cannot be represented
by scales, obviously. Which is to say that Horn scales have their limitations;
but that is not necessarily a bad thing.
Assuming that this is the right way of viewing the connection between
relevance and scales, I will proceed by criticising Horn scales on two points.
In the next section (§3.2) we will see that Horn scales are in a sense too rich.
A Horn scale collects a number of expressions and orders them in terms of
informativeness. I will argue that the ordering is an idle wheel in the machine
at the best of times. That is to say, the generative theory doesn’t need scales,
because sets will do. Though the change from scales to sets is a non-trivial
one, it doesn’t undercut the key tenets of the generative approach. That job
is left to §3.3, where it will be argued that the generative view stands in the
way of a unified theory of Q-implicatures.
54
Chapter 3
3.2 Implicatures and downward entailing environments
Consider the following sentence:
(9)
Fred didn’t read many of the books Wilma gave him.
Assuming, as before, that “many” figures in the scale hsome, many, most, alli,
the substitution method yields (10) as one of the alternatives to (9):
(10) Fred didn’t read all the books Wilma gave him.
Applying the Standard Recipe to this alternative, we get ¬BelS (10) (weak implicature) and possibly BelS (¬(10)) (strong implicature). But this prediction
is patently wrong: since (9) entails (10), it cannot be that ¬BelS (10), let alone
that BelS (¬(10)). On the other hand, (9) seems to license a Q-implicature
that the substitution method does not account for, viz. that BelS (¬(11)); or
in prose: according to the speaker, Fred read at least some of the books Wilma
gave him.
(11) Fred didn’t read any of the books Wilma gave him.
(To avoid irrelevant complications, we’ll assume that “some” and “any” are
just different forms of the same word: one for positive, the other for negative
environments.) Unlike (10), (11) is stronger than (9), and therefore should
qualify as a true alternative, which it isn’t according to the definition in (4).
Note, furthermore, that when (10) is uttered in its own right it may also
implicate that the speaker doesn’t believe (11), despite the fact that “all” is the
strongest element of its scale, and therefore the generative procedure doesn’t
yield any alternatives for (10). Apparently, as things stand the substitution
method produces the wrong alternatives for negative sentences like (9) and
(10).
What causes the problem is that the substitution procedure only checks
for the occurrence of certain telltale words, i.e. scalar expressions like “some”,
“or”, or “clever”, and ignores the remainder of the sentence that contains them.
The procedure presupposes that when we take a statement and substitute a
stronger term for a weaker one, the result is bound to be a stronger statement.
This is not true for negative statements, because negation reverses the direction
of entailment: if ϕ entails ψ, then ¬ψ entails ¬ϕ. Negative environments
are not the only ones to have this property: it is the defining property of
3.2 Implicatures and downward entailing environments
55
downward entailing environments. Downward entailment and its dual, upward
entailment, are defined as follows:
Upward entailing environments
ϕ[...] is an upward entailing environment iff, for any two expressions α
and β: if α entails β, then ϕ[α] entails ϕ[β ].
Downward entailing environments
ϕ[...] is a downward entailing environment iff, for any two expressions α
and β: if α entails β, then ϕ[β ] entails ϕ[α].
Another way of saying this is that an upward entailing environment ϕ[...]
preserves the direction of the entailment relations between the expressions
that can fill the dots; a downward entailing environment, on the other hand,
reverses entailment relations. To illustrate, assuming that “scarlet” entails
“red”, and using “⇒” to symbolise entailment:
(12) Fred is wearing scarlet socks ⇒ Fred is wearing red socks.
Fred isn’t wearing red socks ⇒ Fred isn’t wearing scarlet socks.
Hence, while “Fred is wearing [...] socks” is an upward entailing environment,
“Fred isn’t wearing [...] socks” is a downward entailing environment.
For examples of downward entailing environments that are not negative
(at least not in any intuitive sense), we turn to the quantifiers. In (13)-(15),
square brackets are used to demarcate the restrictor (first argument) and the
nuclear scope (second argument) of the quantifiers “some”, “at most five”, and
“all”. These examples show that “some” is upward entailing in both arguments,
“at most five” is downward entailing in both arguments, and “all” has a mixed
profile; Table 1 gives a summary.
(13) Some of [the students with scarlet socks] [passed]
⇒ Some of [the students with red socks] [passed]
quantifier
some
at most five
all
restrictor
UE
DE
DE
nuclear scope
UE
DE
UE
Table 1: Entailment properties of three quantifiers.
UE = upward entailing; DE = downward entailing.
56
Chapter 3
Some of [the students who passed] [wore scarlet socks]
⇒ Some of [the students who passed] [wore red socks]
(14) At most five of [the students with red socks] [passed]
⇒ At most five of [the students with scarlet socks] [passed]
At most five of [the students who passed] [wore red socks]
⇒ At most five of [the students who passed] [wore scarlet socks]
(15) All [the students with red socks] [passed]
⇒ All [the students with scarlet socks] [passed]
All [the students who passed] [wore scarlet socks]
⇒ All [the students who passed] [wore red socks]
One of the main reasons why entailment properties have received quite a lot
of attention in the linguistic literature is that they affect the distribution of
certain expressions. All languages have negative polarity items, which are socalled because they can occur within the scope of a negative expression, and
are banned from positive environments. English “any” is a case in point:
(16) a. Wilma {*has/doesn’t have} any luck.
b. {*Someone/No one} has any luck.
On closer inspection, it turns out that negative polarity items do not necessarily require a negative environment, though there certainly are constraints
on where they may occur, as the following sentences illustrate:
(17) a. *Some nurses with any experience passed the exam.
b. *Some nurses who passed the exam had any experience.
(18) a. At most five nurses with any experience passed the exam.
b. At most five nurses who passed the exam had any experience.
(19) a. All nurses with any experience passed the exam.
b. *All nurses who passed the exam had any experience.
The generalisation is that negative polarity items may only occur in downward entailing positions. In a sense, a negative polarity item signals that the
environment in which it occurs is downward entailing.3
To return to implicatures, the reason why, despite the fact that “all” is
the strongest quantifier in its scale, (20) still manages to give rise to a Qimplicature is that it happens to occur in a downward entailing environment.
3.
Actually the facts about polarity are rather more complex than this brief discussion may
suggest. See van der Wouden (1997) for an overview of the main issues.
3.2 Implicatures and downward entailing environments
57
(20) Fred didn’t read all the books Wilma gave him. (= (10))
Given that “all” is stronger than any of its scalemates, (20) is weaker than:
(21) Fred didn’t read any of the books Wilma gave him.
Therefore, the weak implicature associated with (20) should be ¬BelS (21),
and the strong one BelS (¬(21)).4
Since negation is not the only environment to be downward entailing, we
should observe this phenomenon in some non-negative environments, as well;
and we do. We have seen that the restrictor of a universal quantifier is a
downward entailing environment. Now consider (22):
(22) Everyone who read all these books passed the exam.
Here the universal quantifier “all” occurs in the restrictor of an “every”-type
quantifier, so we will get a stronger statement if we replace “all” by a weaker
term, like “any”, for example:
(23) Everyone who read any of these books passed the exam.
Therefore, we should expect that an utterance of (22) may implicate that
¬BelS (23), and possibly BelS (¬(23)). These predictions seem to be correct.
Let us now consider how we could amend the substitution method so as to
accommodate the foregoing observations. It seems that what we have to do,
if we want to adhere to the view that alternatives are generated by scales, is
stipulate that scales are reversed in downward entailing environments: if “all”
occurs in a downward entailing environment, its scale isn’t hsome, many, most,
alli, but hall, most, many, somei, and this is what causes (20) to be weaker than
its alternatives with “many” and “some” (Fauconnier 1978, Levinson 2000).
This explanation has one major drawback: it’s wrong. Scale reversal implies that, in a downward entailing environment like (20), “all” becomes weaker
than “some”, “many”, and “most”. But that’s absurd. Surely, the word “all” is
always stronger than its scalemates. So the concept of scale reversal is untenable. True, this notion seems entirely plausible at first, but that’s because it
4.
One problem with this explanation is that, although the following alternatives are also
stronger than (20), it seems that scalar implicatures are less likely to arise here than in the
case of (21).
Fred didn’t read most of the books Wilma gave him.
Fred didn’t read many of the books Wilma gave him.
I don’t know what causes this discrepancy, but it is presumably related to the fact, observed
by Horn (1972), that there is a positive correlation between the robustness of a scalar
implicature and the difference in strength between the scalar expressions involved.
58
Chapter 3
is mistaken for something else. We have seen that there is a sense in which
a downward entailing environment reverses entailment: given that “scarlet”
entails “red”, “not red” entails “not scarlet”; i.e. the direction of entailment
between the complex expressions runs against that between their adjectival
parts. But it patently does not follow from this that in a negative context
“red” entails “scarlet”.
If we really want to stick to the idea that alternatives are generated by
lexical scales, then the best thing to do is give up on the idea that scales are
scales. This must sound a bit Zen, so let me explain. The original proposal was
that the quantifiers “some”, “many”, “most”, and “all” line up in a scale hsome,
many, most, alli. But that’s what they do anyway: it follows from the lexical
meanings of these quantifiers that “many” entails “some”, and so on. We may
as well leave the ordering implicit, and replace the scale with the unordered
set {some, many, most, all}. So instead of assuming that words have scales
associated with them, we now assume that they are associated with sets of
related words, which we might call “Horn sets” (Sauerland 2004). Alternatives
are now generated as follows:
(24) Using Horn sets to generate alternatives
ϕ[β ] is an alternative to ϕ[α] iff α and β share a Horn set and ϕ[β ] is
stronger than ϕ[α].
The crucial change with respect to previous definition in (4) is that the notion of strength does not enter at the lexical level anymore; rather, it is full
sentences that are compared for strength. A sentence has to be weaker than
any of its alternatives, but we’re no longer trying to derive this ordering from
lexical scales; there was no point doing so, anyway, because that is already
being taken care of by the meanings of the sentences in question. Table 2
illustrates the predictions yielded by the amended procedure.
A Horn scale does two things: it collects a number of expressions and
orders them according to their strength. We have seen that there is no work
to do for the ordering imposed by a Horn scale, since it is the entailment
properties of entire sentences, rather than their parts, that is relevant for
defining alternatives. If we want to adopt a generative approach to scalar
implicatures, the only real work for a Horn scale to do is to select a number
of alternative words; which is to say that it doesn’t have to be a scale: a set
will do.
The move from scales to sets may look like a relatively minor modification,
and technically speaking it is. However, the change in spirit is considerable.
3.3 Disjunction: exclusivity and ignorance
i.
ii.
iii.
iv.
v.
vi.
vii.
viii.
59
sentence
alternatives
Wilma read some of the papers
(ii ), (iii ), (iv )
Wilma read many of the papers
(iii ), (iv )
Wilma read most of the papers
(iv )
Wilma read all of the papers
—
Wilma didn’t read any of the papers
—
Wilma didn’t read many of the papers
(v )
Wilma didn’t read most of the papers
(v ), (vi )
Wilma didn’t read all of the papers
(v ), (vi ), (vii )
Table 2: Predictions yielded by the amended procedure
for generating alternatives.
The original intuition underlying the generative view is that the alternatives
involved in scalar implicatures are generated by words, and the previous discussion shows that this intuition is wrong: in the final analysis, it is decided
at sentence level what the alternatives are.
3.3 Disjunction: exclusivity and ignorance
Having argued that the classical Horn scales should be replaced with a notion
that, though similar in one respect, is quite different in another, I will now
try to show that the generative view on scalar implicatures runs into rather
serious difficulties. The argument will proceed by way of the pragmatics of
disjunction.
Disjunctions belong to the stock-in-trade examples of sentences that license
Q-implicatures. In fact, they come with two kinds of Q-implicatures. On the
one hand, disjunctive statements generally give rise to ignorance inferences
(§2.3), while on the other hand, “or” may be given an exclusive construal,
which is standardly analysed as a scalar implicature, and the main concern of
this section. In the following I will try to show that, if the generative view were
right, it would be a mystery how the exclusivity of “or” can be a Q-implicature.
Here’s the example we’ll be working with:
(25) Wilma is dating Albinoni or Boccherini.
Let’s say that the logical form of (25) is “A ∨ B”, where “∨” is the standard
symbol for inclusive disjunction: “A ∨ B” is true iff Wilma is dating Albinoni
60
Chapter 3
or Boccherini or both. Now we can summarise the various pragmatic inferences
associated with (25) as follows:
(26) Ignorance inferences
¬BelS (A) and ¬BelS (¬A)
¬BelS (B) and ¬BelS (¬B)
(27) Exclusivity inferences
[weak ]
¬BelS (A ∧ B)
BelS (¬(A ∧ B)) [strong]
In §2.3, we saw how ignorance inferences can be analysed as Q-implicatures.
Here I would merely like to point out that there is no way the substitution
method could account for these inferences, for the simple reason that the
alternatives involved are shorter than the sentence uttered: it is obviously
impossible to derive “Wilma is dating Albinoni” from (25) by substituting
one word for another.5 Hence, the generative view entails that there is a
stark difference between ignorance inferences and exclusivity inferences: the
substitution method is implicated in the latter but not the former. I believe
this is a serious problem, and will return to it at the end of this section.
Exclusivity inferences are standardly explained as scalar implicatures, on
the basis of the assumption that (28), whose logical form is “A ∧ B”, is one of
the alternatives to (25):
(28) Wilma is dating Albinoni and Boccherini.
Since the speaker chose not to make this statement, we infer that ¬BelS (A
∧ B), and if the Competence Assumption holds, we get BelS (¬(A ∧ B)), as
well, which is to say that, according to the speaker, Wilma is dating Albinoni
and Boccherini, but not both.
So much for the standard story. There are two problems with it: one
general and one specific to the substitution method. The general problem is
that, in the case of disjunction, it is less than obvious why the Competence
Assumption should hold. For, given that (25) typically licenses the ignorance
inferences in (26), it is anything but evident why the hearer should believe that
the speaker believes either that A ∧ B or that ¬(A ∧ B), and therefore it isn’t
obvious at all that, in general, strong exclusivity can be explained in terms of
Q-implicature. It may well be that the exclusive interpretation of “or” is often
5.
This is not quite true: Sauerland (2004) has a trick for deriving alternatives that are, in
effect, shorter than the sentence uttered. However, even by the author’s own admission, it
is a trick.
3.3 Disjunction: exclusivity and ignorance
61
based on considerations of plausibility rather than implicature, and (25) may
be a case in point: since dating is usually, though by no means invariably, a
one-on-one affair, it may just be rather likely that Wilma is not dating two
men at the same time. Whenever such considerations apply, they will obviously
preempt any exclusivity implicatures, weak or strong, that might have arisen.
It is an empirical question how frequently this happens, but I will not address
that question here. In the following I will ignore this complication and accept,
if only for argument’s sake, the widely held premiss that exclusivity is due to
a Q-implicature.
The more serious problem is specific to the generative treatment of scalar
implicatures, and has to do with long disjunctions like the following:
(29) Wilma is dating Albinoni, Boccherini, or Corelli.
To see what the problem is, consider that, if (28) is the relevant alternative to
(25), then presumably (30) is the relevant alternative to (29):
(30) Wilma is dating Albinoni, Boccherini, and Corelli.
If the speaker doesn’t believe that (30) is true, then he doesn’t believe that
Wilma is dating each of Albinoni, Boccherini, and Corelli. But this is not
enough to qualify as an exclusive construal of (29), for such a construal should
entail that, according to the speaker, Wilma is dating just one man on the list.
In order to get a better view on the problem, let’s have a closer look at the
grammar of disjunction and conjunction. “And” and “or” are usually treated
as binary operators, but the examples in (29) and (30) suggest that they can
take any number of arguments. The most natural way of accommodating this
observation is by generalising the semantics of the underlying connectives, as
follows:
(31) ∨ϕ1 ...ϕn is true iff at least one of ϕ1 , ..., ϕn is true.
∧ϕ1 ...ϕn is true iff all of ϕ1 , ..., ϕn are true.
Both ∨ and ∧ now combine with an arbitrary number of sentences, and are
interpreted, in effect, as existential and universal quantifiers, respectively. It
will be clear that the familiar binary treatment of the connectives comes out
as a special case of this.
On this analysis, the logical form of (29) is ∨ABC, and (30) is parsed as
∧ABC. Hence, the strongest Q-implicature we are going to have for (29) is that
BelS (¬(∧ABC)), which merely says that, according to the speaker, Wilma is
not dating all the men mentioned in the sentence. This is correct as far as it
goes, but weaker than a truly exclusive construal. The predictions made by
62
Chapter 3
the substitution method may be correct for binary disjunctions, but they are
too weak for longer ones.
As far as I can see, this conclusion is pretty much inevitable, because it
only requires the premiss that, in order to deal with long disjunctions and
conjunctions, “and” and “or” will have be construed as in (31), and I find this
analysis more appealing than any other I can think of. However, it may be
instructive to consider what will happen if we revert to the logicians’ binary orthodoxy, and postulate that underlying (29) and (30) are multiple occurrences
of good-old binary “∨” and “∧”. On this account, the logical form of (29) is
either “(A ∨ B) ∨ C” or “A ∨ (B ∨ C)”. If this return to logical lore is to make
any difference for the issue in hand, we will have to assume, furthermore, that
alternatives are computed, not on the basis of the surface form of a sentence,
but rather on the basis of its logical form. (This is quite a departure from the
original intuition underlying the substitution method, but we’ll let it pass.)
Then, if (29) is parsed as “(A ∨ B) ∨ C”, its alternatives are:
(32) a. (A ∨ B) ∧ C
b. (A ∧ B) ∨ C
c. (A ∧ B) ∧ C
Whereas if the same sentence is parsed as “A ∨ (B ∨ C)”, we get the following
alternatives:
(33) a. A ∨ (B ∧ C)
b. A ∧ (B ∨ C)
c. A ∧ (B ∧ C)
Consequently, we will have different sets of scalar implicatures, depending
on how we resolve the scope ambiguity hypothesised for (29). To see where
things go wrong, consider the following table, which gives all the possible
combinations of truth values for A, B, C:
A
B
C
a b c d e
1 1 1 1 0
1 1 0 0 1
1 0 1 0 1
f
0
1
0
g
0
0
1
h
0
0
0
Note that on an exclusive construal of “or”, the interpretation of (29) should
retain the columns in which just one of A, B, and C is true, and eliminate all
others; that is to say, only the shaded columns should remain. The problem is
that either set of alternatives will eliminate one column too many: on the first
3.3 Disjunction: exclusivity and ignorance
63
parse, the alternative in (32b) will give rise to the implicature that BelS (¬((A
∧ B) ∨ C)), which entails BelS (¬C), thus ruling out column (g); whereas if
we adopt the second parse, the alternative in (33a) will yield the implicature
that BelS (¬(A ∨ (B ∧ C))), which entails BelS (¬A), and therefore rules out
column (d). In the former case, (29) is interpreted as “Wilma is dating Albinoni
or Boccherini, but not both”; in the latter, as “Wilma is dating Boccherini or
Corelli, but not both”. Either way, the construal we end up with is too strong.
To summarise, it seems unlikely that the substitution method can provide
us with an adequate account of the exclusive readings of long disjunctions. On
the most promising analysis of long disjunctions, the readings it delivers are
too weak, and if we revert to a binary analysis of “or”, which is decidedly less
plausible, the predicted readings are too strong.
On reflection, it isn’t surprising that the substitution method should run
into trouble with long disjunctions. For, in order to arrive at an exclusive
interpretation of (29), the following alternatives will have to be excluded:
(34) a.
b.
c.
d.
Wilma
Wilma
Wilma
Wilma
is
is
is
is
dating
dating
dating
dating
Albinoni and Boccherini.
Albinoni and Corelli.
Boccherini and Corelli.
Albinoni, Boccherini, and Corelli.
Once we have these alternatives, the exclusive reading of (29) can be derived
without further ado. The problem is that only (34d) is derivable from (29) by
way of substitution: each of (34a-c) is shorter than (29), and therefore out of
reach of the substitution method.
Having framed the issue thus, I would like to return to the ignorance implicatures associated with “or”. In the case of (29) these would be of the form,
“The speaker doesn’t know whether or not Wilma is dating x ”, where x is
either Albinoni, Boccherini, or Corelli. These implicatures are enabled by the
following alternatives:
(35) a. Wilma is dating Albinoni.
b. Wilma is dating Boccherini.
c. Wilma is dating Corelli.
As I pointed out at the beginning of this section, the substitution method
cannot account for ignorance implicatures because the relevant alternatives
are shorter than the uttered sentence. But this is the same problem that long
disjunctions give rise to, as we have just seen.
64
Chapter 3
At this point, one begins to suspect that the generative approach is blocking
a unified account of exclusivity and ignorance implicatures. And it is. Here
is how such an account might go. The core alternatives associated with (29)
are (35a-c), but these are not the only alternatives, if we assume (as seems
plausible) that sets of alternatives are closed under conjunction: if ϕ and
ψ are alternatives, then “ϕ and ψ” is an alternative, as well. Therefore, in
addition to (35a-c), the sentence’s alternatives also include (34a-d). Now the
Standard Recipe applies to these alternatives, as always, yielding exclusivity
inferences based on the alternatives in (34) and ignorance inferences based on
the alternatives in (35). A unified explanation if there ever was one.
To be sure, this story is not quite without its problems. For example, it
is natural to suppose that, if sets of alternatives are closed under conjunction,
they should be closed under disjunction, as well: if ϕ and ψ are alternatives,
then “ϕ or ψ” should be an alternative, too. But this we can’t allow, because it
will get us into the same kind of trouble the substitution method runs into if it
adopts a binary analysis of the connectives. However, although it is certainly
a cause for concern, I will not address this issue here, because it is only the
tip of an iceberg that a large part of Chapter 6 will be devoted to.
I have argued that although the generative approach has no problems explaining the exclusive construal of binary “or”, the proposed explanation fails
for longer disjunctions. Moreover, even in the binary case, the generative view
forces us to adopt a strict dichotomy between exclusivity and ignorance inferences, which become amenable to a unified treatment only by giving up the
presupposition that exclusive “or” derives from alternatives generated by the
hor, andi scale. More generally, the view that (some) alternatives are generated
from Horn scales (or sets) yields a distorted picture of Q-implicatures.
3.4 Conclusion
Frankly, I’m a bit wary of the notion of scalar implicature, because it seems to
me that all too often it is intimately connected with the generative view, which
I’ve been criticising in the foregoing. If scalar implicatures are defined as those
Q-implicatures which derive from alternatives generated by the substitution
method, then I’m against them. However, this is not to say that I don’t
believe in scalar implicatures tout court. My point is just that from its very
inception the notion of scalar implicature has been associated with the wrong
sort of theory. My main objection against that theory is that it stands in
3.4 Conclusion
65
the way of a truly unified treatment of Q-implicatures. In the foregoing, I
illustrated this with the two types of Q-implicatures associated with “or”, but
my point is more general: if we adopt the substitution method for dealing
with scalar implicatures, we create an undesirable dichotomy between scalar
and non-scalar Q-implicatures.
Rather than treating scales as primitives and using them to generate alternatives, as the generative approach tries to do, I believe it is better to
take alternatives as primitive, and start from there. If we do so, we can still
distinguish between Q-implicatures according to the alternatives they derive
from, but such distinctions need not be as profound as what is entailed by the
generative approach. A little reflection shows that, lengthwise, there are three
ways a sentence can relate to any of its alternatives: it can be shorter, the
same length, or longer:
(36) a. Betty read “Ulysses” or “Finnegans Wake”.
b. Betty read “Ulysses”.
(37) a. Betty read some of these novels.
b. Betty read all of these novels.
(38) a. Betty read “Finnegans Wake”.
b. Betty read “Finnegans Wake” and “Ulysses”.
Each of the (a) sentences can license the implicature that the speaker doesn’t
believe that the corresponding (b) sentence is true, and while (37a) and (37b)
are roughly the same length, (36a) is longer than (36b), and (38a) is shorter
than (38b). These differences may be important, for if an alternative ϕ is
significantly longer than the sentence the speaker chose to utter, one of the
reasons why ϕ was not uttered could be that it wouldn’t have been worth
the extra effort, in which case the Standard Recipe wouldn’t apply (cf. Horn
2000). This being so, the Q-implicature associated with (38a) should be less
robust than the ones associated with (36a) or (37a), which seems to be right.
See §6.6 for further discussion of this topic.
If we had to define scalar implicatures in purely descriptive terms, we might
say that a scalar implicature is a Q-implicature associated with a sentence ϕ[α],
where the relevant alternative is of the form ϕ[β ] and α and β are of roughly
the same length, as a consequence of which ϕ[α] and ϕ[β ] are similar in length,
as well. It seems to me that, implicitly or explicitly, many scale theorists would
agree with this description, and for the reason given in the last paragraph I
believe that scalar implicatures, thus characterised, are an interesting special
case of Q-implicature.
Psychological plausibility
4
For Grice, conversational implicature is a chapter in what he calls “philosophical psychology”. I will not attempt to define what he means by this, but it is
obvious enough that Grice’s theory is deeply imbued with such psychological
notions as belief, intention, and so forth. It is somewhat remarkable, therefore,
that the Gricean framework is often claimed not to be a psychological one, or
not to make “psychological predictions”. In a sense, this is true. When taken
on its own, Grice’s general theory of conversational implicature makes no predictions about what will be observed in everyday life or in the psychology lab,
and the Gricean theory of Q-implicatures outlined in the foregoing is in the
same boat. For example, though we predict that, under such-and-such circumstances, an utterance of sentence ϕ will have ψ as an implicature, this does
not entail that, under said circumstances, hearers will interpret ϕ as implying
ψ. A Gricean theory of implicature is about inferences hearers are entitled to
make, not about the inferences they will make.
Nevertheless, it is a mistake to write off the Gricean approach as an idle
exercise in philosophy devoid of any bearing on “real” psychology whatsoever.
For, given its distinctively psychologistic signature, it is only a small step from
a Gricean theory of pragmatics to a theory of experimental pragmatics. All it
takes is minimal auxiliary assumptions like the following, for instance:
i. If, according to the Gricean theory, ϕ implicates that ψ, then hearers will
tend to derive ψ as an implicature.
Or alternatively:
ii. If, according to the Gricean theory, ϕ implicates that ψ, and hearer H
actually infers ψ from ϕ, then H has derived ψ as an implicature.
4.1 Charges of psychological inadequacy
67
(i ) is a much stronger claim than (ii ), and therefore less likely to be true. But
both illustrate my point equally well, which is that it doesn’t take that much
to make the Gricean theory yield an ample supply of predictions that can be
tested in the lab.
So, it’s easy enough to interpret Gricean pragmatics as a psychological
theory. But is it a good theory? My answer to this question is an unqualified
yes. It is often said that the Gricean theory of conversational implicature is
implausible from a psychological point of view. Even self-confessed Griceans
have been careful to dissociate themselves from claims to the effect that Grice’s
work is a sound basis for a psychological model of pragmatic inference. I’m
not one of them. I believe the Gricean approach is eminently plausible in every
sort of way, including the psychological.
The purpose of this chapter is twofold. To begin with, I will try to dispel
the myth that the Gricean framework is inadequate from a psychological point
of view. However, what starts out as defensive enterprise will gradually turn
into a more positive argument, as we will be surveying all sorts of experimental
data that favour a Gricean approach to experimental pragmatics.
4.1 Charges of psychological inadequacy
Unfortunately, the view that Gricean pragmatics is psychologically implausible
has never been argued for in any detail. The following is rather typical: commenting on the Gricean way of deriving conversational implicatures, Wilson
(2000: 416) writes:1
It is hard to imagine even adults going through such lengthy chains of inference ...
And that’s all Wilson has to say about this topic. There may be the rudiments
of an argument here (to which we will return shortly), but it is clear that
Wilson is voicing a sentiment more than anything else. The following passage
is atypical in that it spends more than two lines dismissing the Gricean account
of Q-implicature:
In particular, its implications for processing are less attractive. According to
such an account, the inference from the utterance to its scalar implicature
goes through a consideration not just of what the speaker said and the context
but also of what the speaker might have said but did not. It is this type of
1.
The following statement is reiterated (verbatim) by Sperber and Wilson (2002: 8-9).
68
Chapter 4
onerous inference that makes the Gricean account of implicature derivation
seem implausible from a cognitive and developmental point of view. (Noveck
and Sperber 2007)
Though less succinct than Wilson’s statement, this is hardly more explicit.
Like Wilson, Noveck and Sperber merely hint at a possible argument without
developing it in any detail.
Basically, some scholars just seem to experience a powerful intuition that
the Gricean story makes pragmatic inference too difficult for ordinary mortals
to handle. What might cause this feeling? There are five possible reasons I
can think of. Looking at what I have called the Standard Recipe for deriving
Q-implicatures, I see the following ways in which it might be thought to be
implausible from a psychological point of view:
i. Logical complexity: The logical structure of the argument is too complex.
I take it that this is Wilson’s worry.
ii. Abduction: Conversational implicatures are abductive inferences, and abduction is hard.
iii. Incremental processing: According to Grice, you have to know the meaning of a sentence before you can start deriving conversational implicatures.
This premiss sits uneasily with all the experimental evidence showing that
utterances are interpreted incrementally, i.e. word by word.
iv. The intentional stance: The derivation of a conversational implicature
involves reasoning about the speaker’s intentional state: what he wants,
believes, and so on. Such reasoning is hard, especially for children, and
it is improbable that hearers are doing this all the time, as they should if
the Gricean view is to be upheld.
v. Alternatives: It is psychologically implausible that hearers are constantly
monitoring what the speaker could have said, but didn’t. This is Noveck
and Sperber’s point.
In the following, we will consider each of these arguments in turn. The first
three will be disposed of rather swiftly, but I will dwell longer on the fourth one
and longer still on the fifth. This differential attention reflects my belief that
the last two objections are more serious than the others. It is the essence of
Grice’s view on interpretation that speakers take the intentional stance towards
each other, and should this assumption prove to be wrong, then that’s the end
of the Gricean story. The last objection is directed specifically against the
Gricean analysis of Q-implicatures: if there are good empirical reasons for
4.2 Logical complexity
69
doubting that hearers take into consideration not only what the speaker says
but also what he could have said, then that analysis is of philosophical interest,
at best, and irrelevant to the psychology of interpretation.
4.2 Logical complexity
Logical reasoning is one of the oldest topics in cognitive psychology, and there
is a wealth of experimental data on how people reason with conditionals, syllogisms, and so on. None of these data indicate that logical reasoning is
intrinsically difficult: some argument forms are easy, some are hard. For example, subjects’ performance on syllogistic reasoning ranges from near-perfect
to abysmal, and the same goes for propositional reasoning: modus ponens is
easy, modus tollens is harder, and so on.2 To illustrate, one of the very easy
argument forms is the so-called “Disjunctive Syllogism”, of which the following
is an instance:
(1)
¬ψ
Fred is not in Antwerp
ψ∨χ
Fred is in Antwerp or Brussels
Fred is in Brussels
χ
Intuitively, it is pretty evident that this is easy, and psychological experiments
confirm that intuition. For example, in a series of studies reported by Braine
et al. (1995), participants’ performance on Disjunctive Syllogism was essentially perfect, ranging as it did between 94% and 100% correct.
It so happens that the logical form of the Standard Recipe is the same as
that of the argument in (1):
reasoning step
logical form
i. Since S didn’t say ϕ, ¬BelS (ϕ)
¬ψ
ii. BelS (ϕ) ∨ BelS (¬ϕ)
ψ∨χ
iii. BelS (¬ϕ)
χ
(abduction)
(competence)
(from (i ) and (ii ))
How hard can this be? It starts with an abductive inference (more about which
anon), and then uses Disjunctive Syllogism to get to its conclusion. Contrary
to first appearances, the logical structure of the Standard Recipe is perfectly
simple.
2.
See Evans et al. (1993) for a general overview, and Geurts (2003) and Geurts and van der
Slik (2005) for reasoning with non-syllogistic quantifiers.
70
Chapter 4
I suspect that authors who, like Wilson, have complained about the “lengthy
chains of inference” required by Gricean reasoning never really tried to go beyond the first appearances. If we are presented with arguments that run to
half a page or more, just to explain that “some” implies “not all”, it may seem
absurd to suppose that this is how we actually do it. If this is what prompted
Wilson’s complaint, she is mistaken. For one thing, the way an argument is
displayed on the page need not reflect how it is implemented by the brain.
Verbal descriptions are necessarily sequential, for instance, while the brain is
a parallel computer. Therefore, we should be careful not to take such descriptions too literally when considering the psychological plausibility of a theory
of interpretation (or any other theory, for that matter). Another problem
with Wilson’s argument is that it is really too impressionistic. Consider visual
perception. What could be easier than seeing a blue train, for example? Practically everybody with normal or corrected-to-normal vision can do it. Even
very young children do it with ease. But then textbooks on visual perception
invariably run into the hundreds of pages. If these textbooks were right, visual
perception should be bafflingly complex. Isn’t that enough to prove that they
are wrong? Evidently not. What it proves is just that we’re unaware of the
perceptual feats we perform throughout our waking lives.
When it comes to assessing complexity from a psychological point of view,
general impressions count for naught. First, because the complexity of mental
processes is not always open to introspection. Secondly, because processes
that are easy for us may require descriptions that make them appear much
harder. Q-implicatures, and conversational implicatures generally, may well
be like this.
4.3 Abduction
In a sense, it is true that abduction is hard. It is form of defeasible inference,
and non-trivial defeasible logics tend to be computationally intractable. But
even if abduction is hard for logicians and computers, it need not be hard
for us. After all, we’re neither computers nor logical proof systems, and it
is plain that in practice most instances of abduction are child’s play. Here
are a few examples. Upon entering the kitchen, you discover that the light
doesn’t come on. What to do? More to the point: what not to do? There
are infinitely many things you don’t do. For example, you don’t call your
electricity supplier to ask if they changed the voltage, or consider the possibility
4.4 Incremental processing
71
that some fundamental law of physics has been abolished. Instead, you check
the light bulb. Second example: You see a woman entering a bank, and ask
yourself how to explain her behaviour. Again, there are indefinitely many
candidate explanations, but the most likely one is that she intends to conduct
some monetary transaction or other. Last example: You’re on a bus driving
into town, and the driver announces: “The next stop is Gare de Lyon.” Why?
The most likely explanation by far is that the next stop is Gare de Lyon, and
the driver wants to apprise his passengers of the fact. Any other explanation is
much less plausible, and most candidate explanations are wildly improbable.
These are only some of the instances of abductive reasoning that we engage
in almost continuously, and it is plain that they are not difficult at all. Indeed,
most of them are “irrepressible”, to use a phrase of Peirce’s (1931). True,
we routinely rely on quite a lot of aductive reasoning that is hard, like the
diagnostic inferences made for us by doctors, mechanics, and metereologists.
But that only shows that it is with abduction as it is with deduction: most
of our abductive inferences easy, some are more demanding, and some require
special training. In all likelihood, conversational implicatures belong to the
easier kind. After all, conversational implicatures are the outcome of our
trying to make sense of other people’s (linguistic) behaviour, and as a species
we don’t seem to have problems with that sort of thing (§4.5 will belabour this
point). If Gricean reasoning was like diagnosing a brain tumour or a blown
head gasket, it would be hard, but fortunately it isn’t like that at all.
4.4 Incremental processing
This worry is noticeable especially among researchers working on language
processing (e.g., Storto and Tanenhaus 2005, Sedivy 2007), and is prompted
by the fact that, according to Grice, conversational implicatures are derived
on the basis of what is said, which is taken to imply that implicatures cannot
be computed before a sentence is completed: “If Grice is right [...] you need
to know the literal meaning or sense of a sentence before you can calculate
its implicatures in a context [...].” (Levinson 1983: 117) What is worrying
about this is that, as we will see in greater detail in §4.6, there is compelling
evidence, introspective as well as experimental, that interpretation is a highly
incremental process: we construct the meaning of an utterance on the fly, as
the words come in, and it would be wholly remarkable if there were a special
class of inferences that could not, as a matter of principle, be drawn before
the end of the sentence.
72
Chapter 4
Oddly enough, this worry does not seem to extend to entailments, which
should be in the same boat as conversational implicatures, since they, too, can
only be drawn from complete sentences. For example, in order to infer from
(2) that none of the boys are snoring, we need the entire sentence, obviously:
(2)
Only the girls are snoring.
However, it will be equally obvious that we can begin to compute entailments
long before the speaker has finished his sentence. If an utterance starts with
“Only the girls ...”, we can already infer that whatever predicate is going to
fill the dots will not apply to any non-girls. Furthermore, in many contexts it
will be pretty clear in advance how a sentence is going to end. If (2) serves
to answer the question “Who is snoring?”, for instance, it is likely (though not
certain) that “Only the girls ...” will be followed by “... are snoring”, and if
he is sufficiently confident that this is how it will end, the hearer can safely
derive a full-fledged entailment in mid-sentence.
If entailments can be derived incrementally, the same goes for conversational implicatures. Consider again a context in which one speaker has just
raised the question who is snoring, and now another proffers the following
answer:
(3)
Some of the girls are snoring.
Clearly, in the given context, by the time the speaker has completed the subject
phrase of (3), her audience can begin to derive the implicature that not all the
girls are snoring. Similarly, if B asks “What are the boys doing?”, and A starts
answering as follows,
(4)
Some of them ...
B can begin to derive an implicature of the form “Not all the boys ...”, without
having to wait for A to finish her answer.
Hence, there are two factors that dissipate the apparent tension between
conversational implicatures and incremental processing. One is that, often
enough, the hearer will be able to guess how a sentence is going to end long
before the speaker has completed its delivery. Moreover, even if it is not yet
clear how the speaker is going to finish, the hearer may often begin to calculate
partial implicatures which are then fleshed out as the speaker proceeds.
While it is true that a full-fledged conversational implicature requires a
full-fledged sentence meaning, this doesn’t mean that hearers can’t process
implicatures on the fly, just as there is no reason why entailments can’t be
computed incrementally. However, just because it is possible, it doesn’t mean
4.5 The intentional stance
73
that this is what hearers generally do. It has been known since Mitchell and
Green (1978) and Just and Carpenter (1980) that the process of interpretation
slows down at sentence boundaries: the so-called “sentence wrap-up”. Though
it remains rather unclear what hearers are doing during these decelerations, it
is widely agreed that wrap-up tasks claim a relatively large share of processing
effort, which indicates that they are important, and that they are devoted to
global rather than local aspects of interpretation. Hence, while it is unlikely
that lexical lookup will be deferred until the end of the sentence, it is entirely
possible that hearers prefer to compute conversational implicatures (or entailments, for that matter) at the wrap-up stage. It remains to be seen whether
or not this is so, but given that a significant part of the interpretation process
takes place at the sentence boundary, it would make a lot of practical sense if
this was the general strategy.
To sum up, the main take-home points of this section are the following.
First, contrary to what has sometimes been suggested, from the fact that
conversational implicatures presuppose sentence meaning it doesn’t follow that
they couldn’t be computed incrementally. Secondly, it is quite possible, and
consistent with the available data, that implicatures are preferably computed
at the wrap-up stage; experimental studies will have to decide whether or
not there is such a preference. But be this as it may, the rumour that the
very concept of conversational implicature is at odds with the evidence on
incremental processing is just that: a rumour.
4.5 The intentional stance
One of the characteristic features of Gricean theories of interpretation is the
presupposition that interlocutors are constantly trying to guess what is on
each other’s minds. In Dennett’s (1987) phrase, they require that the hearer
routinely adopt the “intentional stance” towards the speaker, trying to figure
out what he believes, wants, and so on. Until fairly recently, it was widely
accepted that taking the intentional stance is inherently difficult, that it is a
skill which takes considerable time to mature, and that it may never become
routine even for adults. If this were true, it would seem to imply that the
Gricean theory, though perhaps acceptable as an idealised picture, shouldn’t
be taken too literally.3
3.
See Breheny (2006) for an excellent discussion of the various positions. The notion that
we can take the intentional stance towards other beings, be they human or not, is often
74
Chapter 4
It may be that, pre-theoretically, it is not implausible that the intentional
stance should be difficult to apply, and this intuitive idea gained support from
early experiments with the famous “false belief task” pioneered by Wimmer
and Perner (1983), who employed the following scenario. A little Austrian
boy named “Maxi” puts a chocolate bar into a blue cupboard, and then exits
the scene. In his absence, Maxi’s mother replaces the chocolate from the blue
cupboard to a green one. Now the critical question is, “Where will Maxi look
for the chocolate when he gets back?” Since it is essential to the concept
of belief that beliefs can be out of step with reality, an accomplished mind
reader should know that Maxi will be heading for the blue cupboard, where
he put the chocolate before leaving. By this criterion, Wimmer and Perner’s
child subjects were not accomplished mind readers: in the two studies they
reported, none of the 3-4-year olds and a mere 57% of the 4-6-year olds pointed
at the right cupboard, and it was only the 6-9-year-old children who showed
no great difficulties with the task (86% correct).
For some time, Wimmer and Perner’s findings were widely interpreted
as showing that young children lack a mature concept of belief, and on the
strength of Baron-Cohen et al.’s (1985) findings it seemed evident that autistic
patients tend to suffer from the same handicap. In the meantime, however,
the general mood has clearly changed. To be sure, it was never clear-cut
that failure on the false-belief task has anything to do with belief (or the
concept of belief) per se, since there are other reasons why subjects may fail
this task. It may be, for example, that young children master the concept of
belief, but lack the linguistic competence to report on what they know, or that
the false-belief task is too complex for other reasons. These suspicions were
vindicated by several studies reported over the years, culminating in Onishi
and Baillargeon’s (2005) demonstration that 15-month-old infants can pass
a non-verbal version of the false-belief task (cf. also Southgate et al. 2007,
Surian et al. 2007).
Over the past few decades, it has become increasingly probable that people
are constantly reasoning about each other’s intentional states, that they are
very good at the game, and that children start playing it at a very early age. I
will quickly go through a handful studies to illustrate how early children start
developing these skills:
expressed by saying that we have a “Theory of Mind”. I consider this extremely poor
terminology, and will therefore avoid it.
4.5 The intentional stance
75
• When they are about nine months of age, babies begin following their parents’ line of regard, monitoring what they are attending to. For example,
if an adult introduces a new word to name an object in front of him, the
child will realise that the new word applies to the object the adult is looking
at, not the one the child happens to be looking at himself (Baldwin 1991,
1993).
• Furthermore, there is evidence that even at a tender age, attention following
is a quite sophisticated art. For example, 12-months-old babies follow the
“gaze” of a faceless robot, which is taken to project off the front end, but
they will not do so if the robot fails to interact with them in a meaningful
way (Johnson et al. 1998).
• It has been shown by Liszkowski et al. (2006, 2008) that 12-month-old infants are able and willing to help an adult who is pretending to have lost a
staple remover (say) by pointing at the “missing object”.
The last example is particularly interesting. Not only does it show that even
before they start speaking, children can assess other people’s needs, it also
proves that already at a very early age, children are capable of the cooperative
behaviour that is essential to Gricean pragmatics.
None of this is to suggest, of course, that infants are accomplished mind
readers. It is generally agreed that mind-reading skills take time to mature,
although there is considerable disagreement over the course of development
and the time it takes. It seems safe to suppose, however, that it will take
at least a number of years before a child fully masters the intentional stance.
If this much is right, we should expect that pre-schoolers are less adept at
deriving conversational implicatures. Since Noveck’s (2001) pioneering study,
which was followed up by many others, we know that this is indeed the case
(Papafragou and Musolino 2003, Guasti et al. 2005, Pouscoulous et al. 2007).
Also pertinent in this connection is Pijnacker et al.’s (2009) finding that autistic patients, who typically have difficulties with the intentional stance, may
derive scalar implicatures at lower rates than normal controls.
To return to our main theme, what can we say about the objection that,
since reasoning about intentions is inherently difficult, it is unlikely that hearers are doing this all the time, as they should if Grice was right? Up to a
point, this objection is valid. It is not unlikely that taking the intentional
stance is hard, as even our closest primate relatives are not nearly as good at
it as we are. But on the other hand, there can be no doubt anymore that our
species has evolved to develop sophisticated mind-reading skills in a matter
76
Chapter 4
of years and that we go on honing these skills in our social interactions. In
short: instead of undermining it, psychological research on the attribution of
intentionality strongly supports the Gricean view.
Before closing this section, I would like to comment on an objection that
is sometimes raised against the notion that speakers are constantly engaged in
reasoning about each other’s intentional states. The objection is that it might
well be that, in reality, people don’t have a particularly deep understanding
of each other’s beliefs and desires, nothing like a genuine “Theory of Mind”
(cf. note 3); they merely seem to be taking the intentional stance. In fact, so
the objection goes, people have an arsenal of quick-and-dirty heuristics that
allow them to simulate the intentional stance. I wouldn’t be too surprised if
something like this turned out to be the case. However, I don’t really see what
the issue is supposed to be about. Even if people are merely pretending to be
taking the intentional stance, there can be no doubt that they are exceptionally
good at it. Now, if you are really good at pretending to be a good driver, you
are a good driver. Similarly, if you are really good at pretending to be a
Gricean reasoner, you are one. Put otherwise, it doesn’t really matter how the
intentional stance is implemented, as long is it is agreed that it is implemented
some way or other. And this much is not controversial anymore.
4.6 Alternatives
It is one thing to assume that hearers reason about the speaker’s beliefs and
desires all the time. It is another thing to claim that, on top of that, hearers
take into account alternative expressions the speaker might have used but
didn’t. According to Noveck and Sperber, it is implausible to suppose that
hearers constantly keep track of what the speaker might have said, and on the
face of it this worry seems reasonable enough. There are so many things a
speaker can say at any given point, how could a hearer ever hope to cut down
this plethora of alternatives to a set of manageable proportions? Doesn’t it
appear highly probable that this would make the hearer’s job far too onerous,
as Noveck and Sperber say? In my view, this is the most serious threat to the
Gricean programme for Q-implicatures, and I will discuss it at some length.
As mentioned already in §4.4, there is a considerable body of experimental evidence showing that hearers have strong expectations about what the
speaker will be saying next, and that these expectations are recalibrated continuously and very quickly. To get an intuitive feel for the remarkable feats
4.6 Alternatives
77
hearers perform non-stop, consider the following passage from Douglas Adams’
“Hitchhiker’s guide to the Galaxy”, from which I have elided an adjective and
five nouns:
At eight o’clock on Thursday morning Arthur didn’t feel very [...]. He woke
up blearily, got up, wandered blearily round his [...], opened a [...], saw a
[...], found his [...], and stomped off to the [...] to wash.
Although it may be hard to guess which exact expressions are missing from
this passage, it will come as no surprise that the first one is “good”, the second
“room”, and the third “window”. But it will be surprising that the fourth
missing word is “bulldozer”. It should be intuitively clear that these judgments
come quite naturally. Experimental data confirm this intuition, and also show
that these judgments are extraordinarily fast. It wouldn’t be unusual for a
hearer to detect the oddity in Adams’ orginal passage halfway through the
ill-fitting word.
In order to make this a bit more precise, let us have a brief look at processing studies making use of event-related brain potentials (ERPs), an average
measure of electroencephalogram (EEG) activity associated with designated
events, that is, in the present context, the occurrence of critical words (like
“bulldozer” in Adams’ text). ERPs provide a direct record of neuronal activity,
with virtually no delay, and therefore they can be used to track various aspects
of comprehension with high temporal accuracy.
An ERP consists of multiple waveforms superimposed on one another. One
of these is the N400, which was first observed by Kutas and Hillyard (1980).
It is a negative-going potential with an onset at around 200 milliseconds and
a peak at approximately 400 milliseconds after the critical event (whence the
N400’s name). Although the precise interpretation of the N400 has not been
settled yet, it may roughly be characterised as a measure of the surprise value
of a given event. For example, Kutas and Hillyard found a strong N400 associated with the last word in “He spread his warm bread with socks”, which is
caused by the fact that “socks” is highly unexpected in this context.
While in Kutas and Hillyard’s experiment, the unexpectedness of the critical word was rather glaring, subsequent studies showed that the N400 registers much milder degrees of surprise, as well. For example, when Hagoort
and Brown (1994) presented subjects with sentences like (5), they found that
“pocket” elicits a stronger N400 than “mouth”, which agrees with the intuition
that the first word, though perfectly coherent, is less expected in this context:
(5)
Jenny put the sweet in her {pocket/mouth} after the lesson.
78
Chapter 4
It will be clear that hearers’ expectations are strongly influenced by the discourse context. In general, a sentence like “The peanut was in love”, would
yield a massive N400 for the last word. But what will happen when this
sentence is embedded in a discourse like the following:
A woman saw a dancing peanut who had a big smile on his face. The peanut
was singing about a girl he had just met. And judging from the song, the
peanut was totally crazy about her. The woman thought it was really cute
to see the peanut singing and dancing like that. The peanut was {salted /in
love}, and by the sound of it, this was definitely mutual. He was seeing a
little almond. (Nieuwland and van Berkum 2006: 1106)
As reported by Nieuwland and van Berkum, in this context it is “salted”, not “in
love”, that elicits an N400. Apparently, the expectations tracked by the N400
take into account contextual information, as well as general world knowledge.
One of the most salient components of the context is the speaker, and van
Berkum et al. (2008) showed that the N400 takes into account information
about the speaker, too. In van Berkum et al.’s study, people listened to sentences whose meanings might be at odds with reasonable assumptions about
the speaker’s sex, age, or social status, as inferred from the speaker’s voice.
For example, when “I have a large tattoo on my back” was spoken in an upperclass accent, there was an obvious discrepancy between social elevation and
dorsal decoration, which became apparent at the critical word “tattoo”. This
type of mismatch, too, turned out to yield an N400.
What these ERP studies show, then, is that hearers check incoming utterances against their expectations practically in real time, word by word. For all
intents and purposes, it is established immediately whether or not any given
word is unexpected at the point at which it occurs. But this is to say that
the hearer must be considering alternative expressions all of the time. Moreover, doing so doesn’t seem to be onerous, at all. If the processes underlying
the N400 were effortful, they should be relatively slow, and they are anything
but. The N400 peaks 400 milliseconds after the onset of the critical word,
which means that, in many cases, unexpected words are recognised as such
well before their delivery is complete.
These results refute Noveck and Sperber’s claim that it is psychologically
implausible to suppose, as the Gricean account of Q-implicature must, that
hearers routinely take into account what the speaker could have said but didn’t.
On the contrary, the ERP literature suggests that this is precisely what hearers
do; if anything, the ERP data argue in favour of the Gricean view.
4.6 Alternatives
79
However, although in general terms ERP data are in agreement with the
Gricean approach, they don’t show that hearers entertain the sort of alternatives that, on the Gricean view, are instrumental in the derivation of Qimplicatures. What we need is experimental evidence qualitatively similar to
that reviewed above, which bears on the question whether hearers routinely
take into account alternatives that are more or less informative than what
the speaker actually said. Such evidence has been provided by Altmann and
Steedman (1988), among others, who presented their subjects with sentences
like (7a,b) in the context of either (6a) or (6b):
(6)
A burglar broke into a bank carrying some dynamite. He planned to
blow open a safe. Once inside, he saw that there was ...
a. a safe with a new lock and a safe with an old lock.
b. a safe with a new lock and a strongbox with an old lock.
(7)
The burglar blew open the safe with ...
a. the dynamite and made off with the loot.
b. the new lock and made off with the loot.
In the context of (6a), (7a) took longer to read than (7b), which is not particularly remarkable, since the first sentence contains a definite noun phrase,
“the safe”, whose intended referent is not resolved by the context. What is
surprising is that in the context of (6b), where both sentences are felicitous,
and had equally long response latencies, the reading times for (7b) were longer
than in the first context. Apparently, participants were slowed down by the
fact that, in the second context, the noun phrase “the safe with the new lock”
was overly informative, or in other words: they expected a shorter and less
informative expression instead.
While Altmann and Steedman’s study shows that hearers monitor the level
of specificity of expressions used by the speaker, Sedivy et al. (1999) report
data suggesting that preferred interpretations are guided, inter alia, by information of this sort. Using an eye movement paradigm, Sedivy et al. presented
their subjects with sentences like (8), while showing them displays in which
there was more than one object matching the adjective. For example, in the
case of (8), the display would contain a tall glass and a tall pitcher:
(8)
Pick up the tall glass.
As in other studies, the subjects in Sedivy et al.’s experiment might look at
either of the two objects upon hearing the adjective, and only fastened on
the target object when hearing the noun. Interestingly, however, Sedivy et al.
80
Chapter 4
showed that a bias for the glass could be induced simply by adding a short
glass to the display. That is to say, if the display contained a short glass in
addition to the tall glass and the pitcher, hearers were more prone to look at
the target glass upon hearing the adjective “tall”.4
The most natural explanation of these findings is a Gricean one: upon
hearing the adjective “tall”, hearers ask themselves why the speaker should
want to use this word, and find a plausible answer in the display before them:
given the presence of a short glass, the adjective might help to select the tall
one. This is Gricean reasoning on a micro-scale, and again it is extremely fast,
since the preference for looking at the tall glass begins to emerge even before
the onset of the noun “glass”.
These effects are Q-based in the sense that, apparently, the participants
in Sedivy et al.’s experiment expected that the critical adjective wasn’t being thrown in for ornamentation, but was meant to serve a definite purpose.
However, following up on the discussion in §1.6, I should like to stress that
I’m not calling this a “Q-implicature”, the reason being not that the kind of
reasoning is qualitatively different (which it isn’t), but rather that its business
is different: whereas Q-implicatures, as the term is commonly understood, are
ampliative inferences deriving from what the speaker has said, the Q-based
reasoning observed in Sedivy et al.’s experiment is involved in determining the
intended referents of definite noun phrases like “the tall glass”. Same sort of
reasoning, different purpose.
The main lesson to be drawn from the experiments reviewed in the foregoing is that, rather than arguing against the notion that hearers routinely track
speakers’ alternatives, processing data suggest that this is precisely what hearers do. The experimental evidence shows that hearers have strong expectations
about what the speaker is going to say next, and that these expectations are
checked and readjusted on a word-by-word basis. Moreover, hearers monitor
the level of specificity of expressions used by the speaker, and use it to guide
their expectations about what the speaker wants to say.
4.7 Conclusion
In this chapter, I have argued that there are no good reasons for believing
that the Gricean theory of implicature is inadequate from a psychological
point of view. Very briefly, the main points were that the logic of Gricean
4.
For further discussion of these and related studies, see Sedivy (2003, 2007).
4.7 Conclusion
81
reasoning, especially the Standard Recipe, is not particularly complex; that
the kind of abductive inferences involved are not particularly hard; that the
experimental evidence on incremental processing is consistent with the Gricean
approach; that it is plausible that interlocutors are always trying to read each
others’ minds; and that it is plausible that, while doing so, they are taking
into account what the other could have said but didn’t.
While the main concern of this chapter was to defuse potential objections
against Gricean theories of interpretation, I believe we can say that the experimental evidence does more than that: it offers strong support for a Gricean
approach to the psychology of interpretation. In particular, Grice’s key insight,
that interpretation crucially involves taking the intentional stance, has been
dramatically vindicated by several decades’ worth of experimental research.
Nonce inferences
or defaults?
5
In §1.4, we reviewed Grice’s distinction between generalised and particularised
implicatures. Grice considers an implicature generalised if it would “normally”,
that is, “in the absence of special circumstances”, be associated with “a certain
form of words” (Grice 1975/1989: 37; see p. 18 for a longer quote). Particularised implicatures, on the other hand, are contingent on specific features of
the context. Though the distinction between generalised and particularised implicatures may have been of strategic importance to Grice, he didn’t see them
as separate species: according to him, both generalised and particularised
implicatures spring from the same pragmatic sources; it is just that, due to
extraneous circumstances, the former happen to arise “normally”. Hence, an
obvious question to ask is why this should happen to be so: why should implicatures be standardly associated, or at least seem to be associated, with some
expressions but not with others? This is one of the questions to be addressed
in the following.
As Grice observed, from generalised implicatures it is but a treacherously
small step to the notion that implicatures may be conventionally associated
with certain expressions.1 If the use of “some” normally gives rise to a “not
all” inference, it is tempting to suppose that “not all” is part of the lexical
content of “some”. Of course, one must be careful not to claim that “not
all” is part of the lexical meaning of “some”, for after all it is cancellable.
But it does seem attractive to asssume that the word “some” comes with
two types of lexical content: a lower-bounding truth-conditional component
1.
If we consider the possibility that implicatures may become conventionally associated with
certain expressions, the question arises what is the exact nature of the association, a question
that has been discussed in a different but related context (i.e., for indirect speech acts) by
Searle (1975) and Morgan (1978), among others. This is a rather intricate discussion, but
as far as I can tell it is not of central importance to our topic of discourse.
83
(“at least some”), which cannot be cancelled, and an upper-bounding nontruth-conditional component (“not all”), which is cancellable, as the following
example reminds us:
(1)
Some of the sheep have the flu. In fact, ...
a. all of them do.
[cancellation of the upper bound ]
b. *none of them do. [attempted cancellation of the lower bound ]
Seen this way, scalar inferences are defaults: “some” will imply “not all” unless
special circumstances indicate otherwise and the inference is cancelled. It
bears emphasising that, on this view, scalar inferences are not conversational
implicatures; rather, they are conventionally associated with certain words. (I
will use “scalar inference” as a neutral term for referring to what I take to be
scalar implicatures.)
Whatever the merits of this line of thinking may be, it should be noted
that it doesn’t apply across the board. It will be evident, for instance, that
inferences like the following are not generated in the lexicon:
(2)
A: What did you have for breakfast?
B: A banana.
; BelB (B only had a banana)
This is readily explained as a Q-implicature, but it surely does not spring
from the lexical content of “a banana”. Hence, the most promising target of
the default view are scalar expressions, and more specifically, the strong scalar
inferences associated with such expressions:
(3)
Some of the sheep have the flu.
; ¬BelS (all the sheep have the flu)
; BelS (¬(all the sheep have the flu))
[weak scalar inference]
[strong scalar inference]
If the word “some” comes pre-packaged with a “not all” inference, the default
interpretation of (3) would imply that, according to the speaker, not all the
sheep have the flu. Hence, the focus of this chapter will be on the strong
upper-bounding inferences associated with scalar expressions, though other
types of Q-implicature will be touched upon, as well.
In the following, I will distinguish two flavours of defaultism, one stronger
than the other, and argue that both are wrong, because scalar inferences are
context dependent to such a degree that they can’t be defaults in either sense.
That is to say, I will argue for a contextualist approach to scalar inferences in
particular, and Q-implicatures in general. However, even if defaultism is on
84
Chapter 5
the wrong track, it has to be conceded that the defaultist view has considerable
intuitive appeal, and one of my objectives in this chapter is to explain what
causes this intuition.
5.1 True defaults
The key concept in this chapter is that of a default inference, and before I start
arguing that scalar inferences aren’t defaults, I want to develop this notion a
bit. The concept of default wasn’t invented by pragmatic theorists. Rather,
they borrowed it from their colleagues in computer science and cognitive psychology. In these disciplines, a default is an assumption which is taken to hold
unless there are specific reasons to suppose that it doesn’t. Defaults, thus
understood, are ubiquitous in everyday life. When you enter a hospital, you
expect to see your typical hospital furniture, as well doctors, nurses, and so
on, although this expectation may be disappointed in infinitely many ways:
it may turn out that the hospital is being renovated, that the building was
converted to a school, or (less likely perhaps) is filled to the roof with chocolate ice cream. When you kick a football, you expect it to fly off, although it
may turn out (to your chagrin) to be filled with chocolate ice cream. And so
on. Such default expectations are mostly unreflected, automatic, and correct
in a sufficiently large number of cases. If every second book was a blank, it
wouldn’t be long before we gave up the default assumption that books contain
pages of print.
Default expectations are essential to everyday reasoning, and since interpretation is to a large extent a matter of everyday reasoning, it will not come
as a surprise that they show up in many processing studies. Here, I will briefly
discuss two experiments by Carreiras et al. (1996), who studied the effects of
gender stereotypes on online interpretation. The gender stereotype associated
with the noun “carpenter”, for example, is that carpenters are male; i.e., carpenters are male by default. Similarly, nurses are stereotypically female.2 In
one of their experiments, Carreiras et al. presented participants with discourses
like the following:
2.
Atlas and Levinson (1981) classify such inferences as conversational implicatures, but since
it is not clear to me how they help to justify the assumption that the speaker is being
cooperative, I’m not convinced that this is right.
5.1 True defaults
(4)
85
a.
Not stereotyped for gender:
The psychologist studied the students’ test results.
{He/She} was concerned about the low standard.
b. Stereotyped for gender:
The babysitter settled down to watch a video.
Then {he/she} heard the baby crying.
Carreiras et al. assumed, on the basis of data collected in a pre-test, that
whereas the noun “psychologist” is not stereotyped for gender, babysitters are
female by default. Therefore, it was expected that, in the first case, reading
times for the feminine and masculine pronouns would be in the same range,
while there should be a delay for “he”, as compared to “she”, in the second
case. In this condition, the noun “babysitter” evokes the default inference that
the person in question will be female, and then the pronoun either confirms
or overrides this expectation, which should make “she” easier to process than
“he”. This is exactly what Carreiras et al. found: while the second sentence
in a discourse like (4b) took significantly longer to read with “he” than with
“she”, no reliable difference between reading times was observed in discourses
like (4a).
This finding shows that it takes effort to overwrite a default assumption
once it has been activated. But what if the noun occurs at a point where
the inference is preempted already? This question was addressed in a second
study, which was conducted in Spanish, and exploited the fact that the Spanish
articles are marked for gender. Here is a sample item:
(5)
{El carpintero/La carpintera} tomó las medidas para hacer el armario.
“The carpenter [masc./fem.] took measurements to make the cupboard.”
In these cases, the carpenter’s gender is established even before the noun is
encountered. Nevertheless, Carreiras et al. still found a strong stereotype
effect: with “la carpintera” the sentence took considerably longer to read than
with “el carpintero”.
Both experiments show that it takes time, and hence effort, to cancel defaults. The second study shows, moreover, that even if the context settles the
matter before the default could possibly arise, it still takes effort to cancel the
default inference.
This is how genuine defaults manifest themselves. They are fast and reflexlike and, as Carreiras et al.’s experiments indicate, they can be quite powerful.
If scalar inferences are lexical defaults, it is only natural to suppose that they
should behave similarly. Therefore, the question arises: Are scalar inferences
anything like this? It will soon become clear that they are not.
86
Chapter 5
5.2 Strong defaultism
I will use the term “strong defaultism” to refer to the position that scalar
inferences are like the gender stereotypes studied by Carreiras et al. On this
view, a scalar inference is triggered by a word and the triggering process is
fast: since “not all” is basically part of the lexical content of “some”, it becomes
available as soon as “some” is retrieved from the mental lexicon.
One problem with the doctrine of strong defaultism is that there has been,
and maybe still is, confusion over who its proponents are. When, around the
turn of the millennium, the first experimental studies on scalar implicature
started to appear, it was said that “neo-Griceans” are committed to strong
defaultism, but such claims are misleading at best: not all neo-Griceans are
strong defaultists, and in particular the author of the neo-Gricean Urtext,
Horn (1972), has never even come near this view.3 The only neo-Gricean who
is committed to strong defaultism is Levinson (2000), though even in his case
it is hard to be completely sure, because Levinson’s views on conversational
implicature are neither univocal nor consistent. However, it is evident that
strong defaultism is one of the main threads in his thinking, for he argues at
length that what he (misleadingly) calls generalised conversational implicatures enhance the efficiency and speed of language processing, which would
make little sense if he thought they are laborious and time consuming. Furthermore, Levinson sees such inferences as
[...] powerful heuristics that give us preferred interpretations without too
much calculation of such matters as speakers’ intentions, encyclopedic knowledge of the domain being talked about, or calculations of others’ mental
processes. (p. 4)
Yet another clue is that Levinson puts scalar inferences on a par with inferences
whose default status is indubitable, like the gender stereotypes discussed in the
previous section, for example. Hence, there can be little doubt that Levinson
is a strong defaultist.
3.
Bott and Noveck (2004: 438) say that what they call “a strong default approach” has been
defended “to some extent” by Horn (1984), implying that Horn’s views are closely related to
Levinson’s. However, although this may have been so in 1984, Levinson went on to espouse
a stronger and blatantly psychologistic version of defaultism, which Horn never subscribed
to. (Bott and Noveck are not overly explicit about what they mean by “a strong default
approach”; the notion is not clearly separated from what I call “weak defaultism”.)
5.2 Strong defaultism
87
Other authors are equally outspoken about their allegiance to the strong
defaultist view:
Our experimental hypothesis amounts to claiming that the implicated content
that is normally associated with uses of the disjunction or does not differ
much from lexical content. Like lexical content, the exclusive meaning of
or should be “closely tied” to the utterance of this lexical item and become
available as soon as the disjunction is heard. (Storto and Tanenhaus 2005:
433)
This quote is interesting not only because of its admirable candour, but also
because it exemplifies the natural affinity between strong defaultism and the
notion that scalar inferences are “closely tied” to certain words. (It isn’t clear
to me why the authors decided to enclose this phrase in quotation marks.)
The main tenet of strong defaultism is about processing. The guiding intuition is that scalar inferences aren’t really inferences; as Storto and Tanenhaus
say, the idea is that scalar inferences are retrieved with the lexical content of
a word; that’s why they are defaults. In the past five years, a series of experimental studies have been conducted to test this theory; in the remainder of
this section, I will discuss what I take to be the most important ones: Bott
and Noveck (2004) and Breheny et al. (2006).4
Bott and Noveck (2004) One of the landmark papers in the experimental
literature on scalar inferences is Bott and Noveck’s (2004). In one of the
experiments reported in that paper, Bott and Noveck presented participants
with sentences like the following, and measured the time participants took to
decide that these sentences were true or false.
(6)
a. Some elephants are mammals.
b. Some mammals are elephants.
c. Some elephants are insects.
(7)
a. All elephants are mammals.
b. All mammals are elephants.
c. All elephants are insects.
The critical item in this experiment was (6a); the others were controls. If a
hearer infers from (6a) that not all elephants are mammals, thus obtaining
4.
Other experimental studies that are relevant here include Bezuidenhout and Cutting (2002),
Noveck and Posada (2003), Feeney et al. (2004), Storto and Tanenhaus (2005), Huang and
Snedeker (2006), Chemla (2007), De Neys and Schaeken (2007), and Pijnacker et al. (2009).
88
Chapter 5
an upper-bounded reading, he should reject the statement. Furthermore, if
scalar inferences are strong defaults, then rejecting the sentence should take
less time than accepting it: according to the default view, a hearer who accepts
the sentence first has to cancel the scalar inference in order to obtain an
interpretation without an upper bound. On the other hand, if scalar inferences
lag behind the computation of truth-conditional meaning, we should see the
opposite pattern: rejecting the sentence should take longer than accepting it.
In the control conditions, Bott and Noveck found that in a clear majority
of cases (over 90% on average), participants gave the expected responses: (6b)
was accepted 89% of the time, (6c) was rejected 93% of the time, and so on.
By contrast, the mean acceptance rate for critical sentences like (6a) was a lot
closer to chance at 41%. Furthermore, most of the participants in this study
(72%, to be exact) failed to be consistent in their responses to the critical
sentences. Obviously, these results sit uneasily with any form of defaultism,
but since our chief concern in this section is with processing, I will not belabour
this point now (see §5.3 for discussion).
Turning to Bott and Noveck’s main finding, they report that the average
“false” response for (6a) was significantly slower than all the other responses
in this experiment, including the average “true” response for the same item.
In all cases except (7a), the difference was well over half a second. Given
that the average response time, which included reading the sentence, was on
the order of 2.5 seconds, this is a considerable delay. Hence, the pattern of
response latencies observed in this experiment disconfirms the claim that scalar
inferences are strong defaults.5
A further experiment by Bott and Noveck used the same materials and
basically the same design. The main difference with the previous study was
that in this version the time available for responding was controlled by the
experimenters, and was either long (3 seconds) or short (900 milliseconds).
This difference didn’t have an effect on the control items, but it did have a
significant effect on the critical items: with the short time lag, participants responded affirmatively in 72% of the cases; with the long time lag, this rate went
down to 56%. Bott and Noveck interpret this result as confirming that scalar
inferences take time to derive, hence as arguing against strong defaultism.
5.
Bott and Noveck’s findings have been replicated in many studies, including Noveck and
Posada (2003), Huang and Snedeker (2006), De Neys and Schaeken (2007), and Pijnacker
et al. (2009). The odd one out is Feeney et al.’s (2004) study, which, for reasons that remain
unclear, failed to reproduce Bott and Noveck’s results.
5.2 Strong defaultism
89
From a logical point of view, Bott and Noveck’s materials contained two
different types of sentences, universal and existential, which between them
gave rise to three types of construal:
All A’s are B’s.
[universal ]
Some A’s are B’s.
[existential without upper bound ]
Some A’s are B’s but not all A’s are B’s. [existential with upper bound ]
Clearly, the upper-bounded construal is more complex than the other two,
but it is important to note that this alone does not suffice to explain Bott and
Noveck’s data. On the strong defaultist view, an upper-bounded interpretation
is the default for all existential sentences, including (6b) and (6c), but these
didn’t slow down participants as the upper-bounded interpretations of (6a)
did. Though it is tempting to assume that this is because (6b) and (6c) are
true (on either interpretation), while the response associated with the upperbounded interpretation of (6a) is “false”, there is no independent evidence that,
in this experiment, positive responses were faster than negative responses. In
fact, with the universal sentences, “true” responses were markedly slower than
“false” responses, although Bott and Noveck don’t report whether or not this
was statistically significant. But anyway, the mere fact that upper-bounded
construals are more complex does not explain the experimental data.
A potential worry about these experiments is the following. Bott and
Noveck asked their participants to decide, for any given sentence, if it was
“true” or not. But what does it mean for a sentence to be true? Many philosophers would say that conversational implicatures have no bearing on the truth
value of a sentence. After all, a conversational implicature is calculated on the
basis of a sentence with a certain meaning, and the truth value of a sentence
is determined by its meaning; any conversational implicatures derived from it
are immaterial as far as its truth value is concerned. We might call this the
“Fregean” view on truth. If the Fregean view is right, then a sentence like (6a)
is true regardless whether a scalar implicature is derived or not.
As a matter of fact, only 41% of Bott and Noveck’s participants said that
(6a) is true, so it was surely not the case that all of them held the Fregean
view on truth. Could it be the case, then, that participants disagreed on the
notion? Could it be that, whereas the “true” responses were given by people
endorsing a Fregean notion of truth, the “false” responses betrayed a stricter
construal, according to which the truth value of a sentence is determined not
only by its meaning but also by such conversational implicatures as it gives
rise to? If this is on the right track, Bott and Noveck’s results tell us very
90
Chapter 5
little about how implicatures are processed. What they show, rather, is that
people disagree on the meaning of the word “true”.
There are at least three reasons for rejecting this explanation. One is that,
as I argued in §1.1, it is doubtful that untutored informants have a Fregean
notion of truth. Secondly, it should be borne in mind that, according to the
view under discussion, scalar inferences aren’t conversational implicatures, but
rather aspects of the sentence’s conventional content. And thirdly, if this
explanation was on the right track, people should become less likely to adopt
the Fregean notion of truth when given more time to think, which seems rather
improbable to me.
All things considered, the explanation proposed by Bott and Noveck is
better than any other I can think of, so I tend to accept that their data argue
against strong defaultism. Nevertheless, it is not an open-and-shut case, for
it can hardly be doubted that upper-bounded construals are more complex
than others, and it may well be that participants’ interpretation of the word
“true” is less straightforward than Bott and Noveck supposed. Still, I don’t
see how these considerations could provide us with an alternative story that
gets strong defaultism off the hook.
Breheny et al. (2006) There is a another concern about Bott and Noveck’s
studies: it is that their design is not very natural. Presumably, the question
whether some elephants are mammals is not very likely to arise in everyday
life. True, experimental conditions are artificial by definition, but the less
so the better. In this respect, Breheny et al.’s (2006) experiments 2 and 3
clearly constitute an improvement; I will review the former here. Consider the
following sentences:
(8)
a. Some of the consultants had a meeting with the director.
b. Only some of the consultants had a meeting with the director.
(9)
a. The director had a meeting with some of the consultants.
b. The director had a meeting with only some of the consultants.
Suppose that “some” in (8) triggers a scalar inference to the effect that not
all of the consultants attended the meeting. Then the content conveyed by
(8a) and (8b) is the same; the only difference is that the upper bound on the
interpretation of “some” is cancellable in the former case but not in the latter,
where it is made explicit by means of “only”. The same, mutatis mutandis, for
(9a) and (9b). Now let an utterance of (8a) or (8b) be followed by:
5.2 Strong defaultism
91
(10) The rest did not manage to attend.
In either case, the anaphor “the rest” refers to the consultants that didn’t
attend the meeting. If scalar inferences are triggered by default, it is reasonable
to expect that this anaphoric link is equally facilitated by (8a) and (8b); for,
on this view, both sentences imply that there were consultants who didn’t
attend the meeting. Again, the same should hold, mutatis mutandis, for (9a)
and (9b), though in these cases establishing an anaphoric link might be faster
than for (8a) and (8b), because here the anaphor is closer to its antecedent. In
a reading time study, Breheny et al. found that these predictions are confirmed
for the sentences in (8) but not for the ones in (9): while in the context of
(8a) and (8b), the reading times for (10) were the same, this sentence took
significantly longer to read when it was preceded by (9a) than by (9b). Hence,
(9a) falsifies the claim that scalar implicatures are strong defaults.
Still, one might be tempted to say, the default view is confirmed by (8a).
However, this cannot be right. If scalar implicatures are defaults, they should
always be triggered: that’s what it means to be default. So how can the
contrast between (8a) and (9a) be accounted for? Breheny et al. propose
the following contextualist explanation. Utterances are always interpreted in
context, even when they are presented “in isolation”, as the sentences in (8) and
(9) were. If a sentence appears out of the blue, hearers will try to construct
a context in which it might plausibly have been uttered, and in doing so they
will take into account such cues as are provided by the sentence itself. For
instance, when a sentence like
(11) The prince was snoring.
is presented in an experimental setting, participants will conjure up a context
in which some prince is part of the common ground between speaker and
hearer.
In (11), it is the definite noun phrase that serves as a cue to what kind of
context the sentence might be uttered in. In the sentences used by Breheny et
al., the relevant cue is part of speech. While in (8a), “some of the consultants”
serves as subject, in (9a) it is the object of the sentence, and since subjecthood
is correlated with topicality, a suitable context for (8a) would be one in which
the consultants are the discourse topic, while (9a) will tend to evoke a context
in which the director is the main character. Put otherwise, the consultants
and their exploits are more relevant in the former context than in the latter,
and therefore (8a) is more likely to give rise to the question whether or not all
consultants were at the meeting, and thus to the implicature that some of the
consultants did not attend.
92
Chapter 5
In this way, a contextualist explanation can be given for the contrast between (8a) and (9a). It may be that the proposed explanation is vaguely
paradoxical, in that it asks us to adopt a contextualist stance on sentences
occurring in isolation. But the air of paradox is dispelled by the observation
that sentences contain context indicators, which provide the hearer (or reader)
with clues as to what sort of context the sentence might be uttered in.
There are two important general lessons to be learned from this special
case. One is that, if an expression α is regularly consorted by an inference
β, it need not be the case that α triggers β by default; only the converse is
true. The second is that, contrary to what one might expect, contextualist
theories do have the wherewithal for explaining data that, prima facie, suggest
that implicatures are defaults. When considered on its own, (8a) could easily
be mistaken for proof that “some” implies “not all” by default. However, the
contrast with (9a) shows that this is not right, and moreover, we can explain
on strictly contextualist principles why there seems to be a default inference
in the case of (8a).
Epitaph on strong defaultism In §5.1, we saw how strongly processes of interpretation can be affected by true defaults (in this case, defaults created
by gender stereotypes). In the meantime, it should have become clear that
scalar inferences aren’t like this at all. Between them, the experimental data
procured by Bott and Noveck, Breheny et al., and others (cf. note 5) render
it highly unlikely that scalar inferences are defaults in the true, strong sense
of the word. By the same token, they argue against theories claiming that the
upper-bounding inferences associated with scalar expressions are hard-wired
into the lexicon. For if they were, they should become available as soon as a
word is retrieved from the lexicon, and experimental data suggest that they
don’t.
5.3 Weak defaultism
Motivating weak defaultism Having dealt with strong defaultism, let us now
turn to weak defaultism, which is the view that scalar expressions give rise
to upper-bounding inferences as a matter of course; it is what will happen
“normally”. This description is admittedly vague, but then the doctrine itself
isn’t always easy to nail down, as the following passage illustrates:
5.3 Weak defaultism
93
By default interpretation, I simply mean the one that most people would
give in circumstances in which the context is unbiased one way or the other.
(Chierchia 2004: 51)
Since every context is “unbiased one way or the other”, I’m not at all sure what
this says, but the author’s intended meaning is clear enough: ceteris paribus,
scalar expressions give rise to upper-bounding inferences.
Unlike its strong cousin, weak defaultism makes no claims about processing.
It merely says that upper-bounding inferences are the norm. But why should
they be? Oddly enough, this question is rarely addressed. More generally,
authors who claim that implicatures are defaults (be they weak or strong)
almost never endeavour to explain why (Levinson 2000 is an exception). Now
in the case of strong defaultism, such an explanation isn’t hard to come by:
strong defaults can be motivated by the assumption that they are encoded in
the lexicon. I have argued that this is wrong, but at least it is an attempt
at explanation. What about weak defaults, then? It is much less obvious
what kind of mechanism might account for weak defaults. I will return to this
question later, and until then treat weak defaultism as a purely descriptive
claim.
What is the evidence for weak defaultism, thus understood? This question,
too, has received less attention than it deserves, but one possible argument is
suggested by the popular notion that implicatures are cancellable:6
(12) a. Betty’s grades are good—in fact, they are excellent.
b. Fred broke an arm or a leg, if not both.
If scalar inferences were defaults, we could say (indeed, we would have to say)
that the second half of (12a) cancels the “not excellent” inference triggered in
the first half, and similarly for (12b). However, sentences like (12a,b) don’t
show that scalar inferences are defaults, because there is no compelling reason
for analysing them in terms of cancellation, in the first place, as I argued in
§1.5. The “if not” construction, for instance, may be construed as signaling
that the speaker is not prepared to commit himself to a stronger claim. Evidence in favour of this view is provided by attested examples like (13a-c), cited
from Horn (1989: 241):
(13) a.
6.
Overt antifeminism, if not homosexuality, may be the result of such
experience in the male.
Although I’m not aware of any explicit versions of this argument in the literature, it seems
to me that it is a natural corollary of the cancellation concept.
94
Chapter 5
b. In the Netherlands the crowds [for the Pope] were small, the welcome
lukewarm if not cold.
c. Most photographers were inarticulate if not subhuman.
Surely, it would be far-fetched to suppose that antifeminism implies heterosexuality by default, so (13a) shows that the “if not” construction doesn’t demand
a default inference for it to cancel; the same, mutatis mutandis, for (13b) and
(13c).
Another argument in favour of granting scalar inferences default status is
given by Levinson (2000). It is based upon certain patterns of lexicalisation
documented in detail by Horn (1989). If we consider the positive duals “all”
and “some”, we see that the negation of the latter is lexicalised while that of
the former is not: there is no single word for expressing “not all”, no word
that stands to “all” as “no(ne)” stands to “some”. The same holds for other
dual pairs, like “necessary/possible”, “obligatory/permissible”, and so on. In
all cases, there is a lexical gap in the same place, as the following diagram
shows:
some : no(ne)
all : [not all]
permissible : forbidden
obligatory : [not obligatory]
sometimes : never
always : [not always]
Horn’s explanation for this pattern, which is found across languages, is that
“some”, “possible”, etc., tend to implicate “not all”, “not necessary”, etc., thus
obviating the need for special lexical entries for the latter group. Levinson
endorses Horn’s proposal, and maintains that it supports the view that scalar
inferences are defaults:
Anyone who rejects [the default view] will have to find another explanation for
this lexicalisation pattern. The [contextualist] cannot explain it—according
to any such theory, there are no general tendencies to be found, or if there
are they have the status of mere behavioral tendencies, playing no role in the
systematic generation of implicatures. (Levinson 2000: 71)
This reasoning is plain wrong (as indeed I pointed out well before Levinson’s
book was published; Geurts 1998b). Horn’s analysis does not require that
scalar inferences be defaults. The keystone of Horn’s explanation is that,
generally speaking, when a speaker should want to use “not all” he can use
the shorter form “some” in its stead, because on such occasions “some” will
tend to implicate “not all”. This, together with the assumption that positive
concepts have a stronger tendency to be lexicalised than negative ones, suffices
5.3 Weak defaultism
95
to explain the lexicalisation pattern observed by Horn. What Levinson wants
to prove is that “some” implies “not all” by default simpliciter ; and this simply
doesn’t follow.
To sum up: thus far, it isn’t clear that there is any explanatory work to
do for the hypothesis that scalar inferences are weak defaults. However, we
haven’t yet considered the possibility that there might experimental evidence
supporting that hypothesis. We’ll go into that soon, but before doing so I
would first like to address the question what provides the intuitive motivation
for weak defaultism.
Defaultist intuitions I suspect that, in the final analysis, the idea that scalar
inferences are defaults owes its popularity to a gut feeling: it just seems intuively plausible that this should be the case. When we ask ourselves whether
we would take an utterance of (14a) to imply that, according to the speaker,
(14b) holds as well, it seems evident that we would:
(14) a. Some of the goats have the flu.
b. Not all of the goats have the flu.
This reasoning seems so compelling that it may be hard to believe that it
could be wrong. Still, the way introspection is employed in this argument is
arguably misleading (Geurts 1998b, 2006, 2009). Obviously, to ask oneself
whether or not (14a) implies (14b) is to suggest already that it might be
implied, but more importantly, this question raises the issue of whether or not
all of the goats have the flu, or in other words, it makes it relevant to establish
whether this is the case. And even if our intuitions about this case are to be
trusted (which is as it may be), they don’t necessarily tell us anything about
how (14a) is interpreted in situations where the issue is not raised. In short,
the introspective evidence used to buttress claims to the effect that scalar
inferences are defaults is dubious, at best.
In the last section, I argued that even if it were the case that a sentence
like (14a) tends to imply (14b), we shouldn’t jump to the conclusion that this
inference is a default triggered by “some”, because there would be a perfectly
good contextualist explanation that doesn’t require defaults. On the contextualist view, since “some of the goats” is the grammatical subject of (14a), the
hearer will try to set up a context for this sentence in which the goats are
the discourse topic, and in such a context it is relevant to ask, apropos of an
utterance of (14a), whether or not all the goats might have the flu. Thus far,
that question is implicit. The point of the last paragraph is that, when we
96
Chapter 5
consult our intuitions about (14a), the question is asked explicitly. Hence, our
intuitions are not about how this sentence would be interpreted in general,
but are confined to a limited range of contexts, that is, contexts in which it is
relevant to establish if all goats have the flu. And if our intuitions tell us that
in such contexts (14a) would imply (14b), that doesn’t further the defaultist
cause in any way.
This argument is supported by the following experiment, which compared
two different methods for probing scalar inferences (Geurts and Pouscoulous
2009). In this experiment, native speakers of Dutch were presented with the
same sentence in two different tasks:
(15) Some of the B’s are in the box on the left.
In one task, the inference condition, participants were asked whether this implies that not all the B’s are in the box on the left. In the verification condition,
participants had to decide whether the same sentence correctly describes the
following situation:
BBBAAA
CCC
Following Bott and Noveck (2004) and others (cf. §5.2), we assumed that
someone who interprets (15) as implicating that not all the B’s are in the box
on the left will deny that the sentence gives a correct description of the facts.
Note that, whereas in the inference task the scalar inference is associated with
a positive response, in the verification task it is associated with a negative
response; therefore we had to control for the possibility of a positive response
bias, which is why we included a liberal number of foils that were superficially
similar to the critical item.
The outcome of the experiment was that scalar inferences were observed
almost twice as often in the inference condition (62%) as in the verification
condition (34%). Furthermore, since participants’ performance on the distractor items in the verification task was nearly perfect (97% correct), it is unlikely
that the low rate of “no” answers for the critical item was caused by a positive response bias. These data confirm that the question of whether or not a
sentence ϕ implies a scalar inference ψ is a biasing one. This is not to suggest
that we should stop asking such questions altogether. Despite their bias, they
remain an invaluable source of information. However, they are useless if we
want to gauge how common these inferences are: the intuition that “some”
normally implies “not all” counts for naught.
5.3 Weak defaultism
97
Lack of experimental support In order to test the claim that scalar inferences
are weak defaults, quantitative data are needed. In particular, we need experiments that probe scalar inferences in a context that is as neutral as possible.
What these experiments should not do is ask informants whether or not a
sentence ϕ is likely to imply another sentence ψ, where ψ spells out a scalar
inference ϕ. As it turns out, the literature contains a fair number of studies that meet these requirements. In fact, we have already encountered two
of them. In the verification tasks employed by Bott and Noveck (2004) and
Geurts and Pouscoulous (2009), people were asked to decide if sentences of
the form “Some A’s are B’s” are true in a situation in which all A’s are B’s.
While in the former study, negative responses were recorded 59% of the time,
the corresponding rate in the latter was 34%. Neither result agrees with defaultist expectations. If defaultism was right “some” should imply “not all” in
the absence of evidence to the contrary, so unless it can be argued that these
experiments provided inadvertent cues that the scalar inferences were to be
cancelled, the negative response rates should be a lot closer to 100% than they
actually were. As far as I can tell, no such cues were provided, and if they
weren’t, the reported data argue against weak defaultism.
Let me give one further example, this time of an experiment on disjunction.
In an acquisition study, Paris (1973) presented children and adult controls with
disjunctive sentences with arbitrary content, like the following:
(16) The bird is in the nest or the shoe is on the foot.
Paris’s materials contained sentences with “or” as well as “either/or”. Participants were asked to determine whether or not such sentences were true of a pair
of pictures. In the critical trials, pictures were shown that made both disjuncts
true. If “or” is construed exclusively, i.e. as implying “not both”, participants
should say that the sentence is false, whereas on an inclusive construal, sans
exclusive inference, the sentence should come out true. Paris’s main finding
was that, overall, inclusive interpretations were preferred for 82% of the sentences with “or” and 76.5% of the sentences with “either/or”. (The difference
between “or” and “either/or” was significant, though much smaller than one
might have expected.) For the adult participants, the rates were 75% and
68.5%, respectively. These data suggest that the normal interpretation of “or”
is inclusive, and therefore go against the defaultist view.
The experiments reviewed in the foregoing are representative of what one
finds in the literature: when contextual factors are factored out and the experimental paradigm is as neutral as possible, rates of scalar inferences are
98
Chapter 5
source
scalar term
Paris (1973)
or
Chevallier et al. (2008), exp. 1
or
Pijnacker et al. (2009)
or
Paris (1973)
either/or
Evans and Newstead (1980), exp. 2
either/or
Braine and Rumain (1981), exp. 3
either/or
Noveck (2001), exp. 3
some
Bott and Noveck (2004), exp. 3
some
Feeney et al. (2004), exp. 1
some
Geurts and Pouscoulous (2009), exp. 2 some
Noveck (2001), exp. 1
might
%
25
25
54
32
33
41
59
59
65
34
65
Table 1: A sample of experimental data on scalar inference. The
rightmost column gives the rates at which upper-bounded interpretations were observed.
below 50%, on average, and never higher than 65%. Table 1 gives a sample of
relevant experimental studies reported since the early 1970s. Clearly, there is
no experimental support for the view that scalar inferences are weak defaults.
5.4 Contextualism
Contextualisation and quasi-defaults I believe we can safely conclude from the
foregoing discussion that there is no empirical evidence that words like “some”
or “or” come with default scalar inferences attached to them. On the contrary,
such evidence as is available militates against this view. Hence, the original
intuition underlying defaultism proves to be mistaken. Even if it should turn
out that the derivation of quantity inferences isn’t entirely dependent on the
context, the truth is a good deal closer to the contextualist end of the table.
Still, I believe it may be argued that, when charitably construed, weak
defaultism may contain a grain of truth, after all, for two reasons. The first
reason has to do with contextualisation; the second, with the possibility of
what I will call “default elements” in the derivation of implicatures. Starting
with the first point, we have seen already that hearers always try to contextualise a sentence, even when it is presented in the “null context”, i.e. without
explicit information as to what kind of context the sentence is supposed to be
5.4 Contextualism
99
uttered in. We have seen that the contextualisation process is guided by various factors, including word order and the presence of a question the sentence
is supposed to answer. These factors have nothing to do with implicature per
se, but since they affect the availability (or salience) of alternatives, they may
have an indirect effect on the derivation of implicatures. Due to such factors,
it may be that a sentence like (17) will by default be interpreted as implying
that, according to the speaker, not all the goats have the flu:
(17) Some of the goats have the flu. (= (14a))
But again, if there is a default inference in this case (which is as it may be),
it is not hard-wired into the lexical meaning of “some” and it doesn’t have
anything to do with implicature as such. A contextualist view on implicature
allows for the possibility that some sentences have a strong tendency to give
rise to certain implicatures.
Let me briefly mention three further factors that, like word order and questions, may influence the derivation of Q-implicatures by raising the availability
of alternatives:
• Focus: It is quite likely that scalar inferences can be facilitated by means of
focus. Compare, for example, (18a) to (18b):
(18) a. The director had a meeting with some of the consultants.
b. The director had a meeting with some of the consultants.
We encountered (18a) in our discussion of Breheny et al.’s experiment (§5.2,
example (9a)), which suggested that this sentence is less likely to give rise
to the scalar inference that the director didn’t meet all the consultants.
However, it seems intuitively plausible that this likelihood will increase when
“some” is highlighted, as in in (18b). This intuition is confirmed in Chevallier
et al.’s (2008) experimental study on disjunction, which shows that focus
on “or” causes higher rates of exclusive interpretations, though it has to
be noted that this is a matter of degree: focus doesn’t force an exclusive
construal, it just makes it more likely. (Chevallier et al.’s study is in line
with Paris’s findings on the difference between “or” and “either/or”, discussed
in the last section.)
Focusing is standardly viewed as a device for evoking alternatives: by
using (18b) the speaker draws attention to the fact that he is claiming that
the director had a meeting with some rather than α of the consultants,
where α would be an expression of the same category as “some”. Naturally,
one of the more obvious candidates for α-hood will be “all”, and if the focus
100
Chapter 5
on “some” is thus interpreted, the likelihood of a “not all” inference is bound
to increase.
• Partitives: Existential determiners like “some” are special in that for them
a scalar construal doesn’t always make sense:
(19) There are some oranges in the fridge.
Normally speaking, it would be odd to infer from this that, according to the
speaker, not all the oranges are in the fridge. Such an interpretation only
can be opportune if the said oranges are part of a larger set, or in other
words, it requires a partitive construal of “oranges”, which in (19) is blocked
by the existential “there” construction. There are various ways of promoting
a partitive construal, some of which we have encountered already: it will
help if the indefinite is the grammatical subject and/or “some” is focused.
Another way is to use a partitive construction, as in (20b):
(20) a. The fridge contains some oranges.
b. The fridge contains some of the oranges.
Whereas in (20a) a partitive construal is comparatively unlikely, in (20b) it is
enforced by the partitive construction “some of the oranges”, and accordingly
this sentence is more liable to give rise to a “not all” inference (cf. Milsark
1977, Horn 1997).
• Relative complexity of alternatives: For reasons discussed in §3.4, it may be
that the likelihood of a Q-implicature is affected by the relative complexity
of the alternatives involved in its derivation. To illustrate, consider the
following type of case:
(21) Fred or Barney made a mistake.
; ¬BelS (Fred made a mistake)
; ¬BelS (Barney made a mistake)
According to my intuitions, these inferences are quite robust, and although I
have been cautioning that such intuitions should not be taken at face value,
I’m fairly confident that this is right. Assuming it is, how could we account
for the robustness of such inferences? The answer, I would like to suggest,
is that in the case of (21) the relevant alternatives are simpler than the
sentence itself. A speaker who utters (21) goes out of his way to deliver a
sentence that takes more effort to produce and interpret than either “Fred
made a mistake” or “Barney made a mistake”, and this is why the ignorance
inferences associated with (21) are so robust. (See §6.6 for further discussion
of this point.)
5.4 Contextualism
101
Default elements in the derivation of implicatures I have suggested a number
of factors that may affect the contextualisation of a sentence, and thus may
conspire to foster the illusion that implicatures are defaults. It should be
emphasised that none of these suggestions imply a concession to defaultism:
everything I have said thus far is consistent with a strictly contextualist stance
on Q-implicatures in general, and scalar implicatures in particular. Still, I’m
about to suggest that some such concessions may be in order. Not that I will
back down from the claim that defaultism is false. As far as I can see, the
notion that Q-implicatures of any kind are default inferences has nothing to
recommend it. However, I am prepared to concede that there may be default
elements in the derivation of Q-implicatures:
• Competence: The notion of competence plays a conspicuous role in the
standard Gricean treatment of Q-implicatures. As discussed in Chapter 2,
Gricean theories first derive a weak implicature of the form ¬BelS (ϕ), which
is strengthened to BelS (¬ϕ) if speaker S is taken to be competent with
respect to ϕ. In §2.3, I suggested, as a working hypothesis, that hearers
adopt the Competence Assumption by default. If this is correct, it doesn’t
alter the case against defaultism in the least. However, it does imply that,
if a Q-implicature arises, it will be strengthened by default. This prediction
seems entirely reasonable to me; whether or not it is correct is an empirical
question.
• Level of specificity: In §2.4, we discussed Matsumoto’s (1995) observation
that, in some cases, Q-implicatures appear to be language specific. One of
Matsumoto’s examples is that, whereas the English sentence (22) does not
normally implicate that the speaker doesn’t know whether Michio is older
or younger than Takashi, its Japanese equivalent does:
(22) Michio is Takashi’s brother.
The contrast between the two languages is obviously related to the circumstance that, besides words corresponding to English “brother” and “sister”,
Japanese also has a system of sibling terms that encode sex as well as seniority. However, this fact alone does not suffice to explain the difference.
What is needed, in addition, is the assumption that, by default, speakers of
Japanese expect each other to use the more specific sibling terms.
Although Matsumoto’s example is especially vivid, I don’t think it is
an isolated case. Speaking more generally, it is plausible to suppose that
hearers will have default expectations about the minimum level of specificity
102
Chapter 5
called for in ordinary situations. Normally speaking, we expect chairs to be
called “chairs” rather than “pieces of furniture” or “things”, dogs to be called
“dogs” rather than “mammals” or “animals”, and so on (cf. Brown 1958, Cruse
1977, Rosch 1978, Hirschberg 1985). Whenever such default expectations
are disappointed, Q-implicatures are bound to ensue, but it is important to
note that it is not these implicatures that have default status, but rather
hearers’ expectations about what a thing would normally be called.
5.5 Conclusion
Prima facie, defaultism seems plausible enough. On reflection, it is obvious
that this view is applicable, if at all, to a ragbag of cases only, which don’t
make up anything like a natural class. It is clear that not all Q-implicatures
could be defaults. Defaultism is prima facie credible when restricted to scalar
inferences, but even so it is easy to come up with instances of scalar inferences that are strongly context-dependent, and experimental evidence suggests rather forcefully that the stock-in-trade licensers of scalar implicature,
like “some” and “or”, do not trigger upper-bounding inferences by default. In
short: defaultism is a lost cause.
None of this is to deny that some implicatures may arise more or less systematically, either in general or (more likely) within a certain range of contexts.
However, as we have seen in the last section, such observations, if correct, are
consistent with the Gricean view that there is nothing about implicatures per
se that should lead us to expect that they, or some of them, have default status. There is simply no explanatory work to do for the hypothesis that (some)
implicatures are defaults.
Intentions, alternatives,
and free choice
6
In Chapter 2, we discussed the Standard Recipe for deriving Q-implicatures,
which is the best-known way of fleshing out the Gricean notion of quantity implicature. In this chapter, we will see that there is another way, as well, which
despite its prima facie similarity to the Standard Recipe, does a much better job of explaining Q-implicatures; I call it the “intention-based” approach.
Intention-based treatments of Q-implicature have been advocated by Robert
van Rooij and Katrin Schulz (both jointly and separately), and Benjamin
Spector,1 and though it may be less popular than the standard account, I
will argue that the intention-based approach is the better choice, mainly because it is more general. While the Standard Recipe works rather well with
scalar implicatures, it becomes unwieldy or worse when applied to other types
of Q-implicature. By contrast, the intention-based approach offers a unified
explanation of all varieties of Q-implicature.
Here’s a first taste of the difficulties we run into when we try to extend the
Standard Recipe beyond scalar implicatures. Suppose George asks Mildred,
“What did you have for lunch?”, and Mildred answers:
(1)
I had some of the strawberries.
Mildred’s answer might yield a scalar implicature to the effect that she didn’t
have all the strawberries. According to the standard view, this implicature is
generated by negating an alternative sentence Mildred could have used in lieu
of (1), namely:
(2)
I had all the strawberries.
So far so good. But there is a further Q-implicature, as well. In the given
scenario, George might well be surprised, apropos of her uttering (1), that
1.
See, e.g., van Rooij and Schulz (2004), Schulz and van Rooij (2006), Schulz (2005), and
Spector (2006).
104
Chapter 6
Mildred had such a light lunch: in this scenario, (1) would be liable to implicate
that Mildred only had strawberries for lunch. Now the question is: What
alternative sentence(s) might this implicature be the negation of? The only
answer, as far as I can see, is that the implicature negates each of a set of
alternatives like the following:
I had
I had
I had
I had
I had
I had
etc.
some
some
some
some
some
some
of
of
of
of
of
of
the
the
the
the
the
the
strawberries and a biscuit.
strawberries and a banana.
strawberries and some muesli.
strawberries, a banana, and a biscuit.
strawberries, a banana, and two biscuits.
strawberries, a banana, and three biscuits.
The chief merit of this solution is that it does its job, but apart from that it is
clumsy and not very plausible from a psychological point of view. The number
of things Mildred could have had for lunch is huge, and if we consider all
possible combinations the number becomes astronomical. It is hard to believe
that George, or any other hearer for that matter, would have to entertain all
these possibilities merely to infer that Mildred had a light lunch.
The intention-based approach offers an alternative way of deriving Qimplicatures that avoids this problem. The idea is really very simple. Whereas
according to the Standard Recipe, George begins by asking why Mildred failed
to produce one of a set of stronger alternative statements available to her, the
intention-based method has him pose the following question:
If Mildred believed (knew) that her lunch comprised more than just some
strawberries, would she have said so?
If George’s answer to this question is “yes”, then he will infer that Mildred’s
lunch contained nothing more than some strawberries. If this is how George
reasons, he need not have any particular alternatives in mind. He merely has
to decide if a hypothetical belief of Mildred’s would have prompted her to
deliver a different statement.
In the following, I will develop the intention-based view using so-called
“free choice” inferences as my main exhibit. The reason for this is partly that
free choice is an interesting and much-debated topic in its own right, and
partly that it causes a serious predicament for the Standard Recipe, while the
intention-based approach explains it without further ado. Or so I will argue.
6.1 Free choice
105
6.1 Free choice
My son asks, “What’s for dessert?”, and I say:
(3)
You can have fruit or chocolate cake.
Bright kid that he is, my son swiftly arrives at the conclusion that he can have
chocolate cake. But what justifies his inference? The proposition that x has
fruit or chocolate cake is entailed by, hence weaker than, the proposition that x
has chocolate cake, and in this sense the permission to have fruit or chocolate
cake grants less than the permission to have chocolate cake. It seems that my
son derived his conclusion from a premiss that is logically weaker, and that
therefore his reasoning can’t be valid. In a nutshell, this is the problem of “free
choice permission” (von Wright 1968, Kamp 1973).2
It bears emphasising, perhaps, that from a semantical point of view, (3) is
a weaker statement than either “You can have fruit” or “You can have chocolate
cake”. (3) is a possibility statement, and possibility is closed under entailment:
if ϕ entails ψ and ϕ is possible, then ψ is possible, too. Assuming, as always,
that a disjunctive proposition is entailed by each of its disjuncts, it follows
that “It’s possible that ϕ” entails “It’s possible that ϕ or ψ” (as well as “It’s
possible that ψ or ϕ”), but not the other way round. The problem of free
choice permission is precisely to explain why it appears to be the other way
round.
Note that, apart from the free choice inferences it gives rise to, (3) would
probably also be construed as implying that the addressee is not allowed to
2.
Some authors, including Zimmermann (2000) and Geurts (2005), have assumed that the
underlying form of (3) is more like (i), and taken the analysis from there:
(i)
You can have fruit or you can have chocolate cake.
There has been some discussion over whether this move is legitimate from a syntactical
point of view (Kamp 1979, Jennings 1994, Geurts 2005), but even if it was, it wouldn’t
suffice to solve the problem. Having expanded (3) into (i), Zimmermann and Geurts have
to resort to a modal analysis of disjunction in order to obtain free choice inferences, which
is not without its problems. Furthermore, it should be noted that a periphrastic expansion
is not always available. For instance, (6) below is obviously not synonymous with:
(ii)
Many people had pizza or many people had pasta.
The account of free choice to be presented in this chapter does not apply to cases like (i).
I still feel that the disjunction in this type of sentence has a non-Boolean flavour, but have
ceased to believe that they are the proper starting point for a general theory of free choice.
106
Chapter 6
have fruit and chocolate cake. However, this exclusivity inference is entirely
unproblematic (the relevant alternative is “You can have fruit and chocolate
cake”) and orthogonal to the problem under discussion. That is to say, free
choice is a problem no matter whether or the sentence is read exclusively or
inclusively. Therefore, I will ignore this aspect in the following.
Although the paradigm cases of free choice inference involve performative
(i.e., permission giving) uses of deontic statements, as in the opening example,
it is important to note that they range more widely. To begin with, reportative
uses of deontic statements license free choice inferences, as well. If I utter
(3), I may be giving my son permission to have fruit or chocolate cake, or
alternatively, I may be reporting on what the restaurant’s menu has to offer,
for example. Free choice inferences arise in the latter case just as they do
in the former. Secondly, we observe free choice inferences with non-deontic
modals:
(4)
George may be British or Canadian.
; He may be British.
; He may be Canadian.
(5)
Betty can balance a fishing rod on her nose or chin.
; She can balance a fishing rod on her nose.
; She can balance a fishing rod on her chin.
The most natural construal of (4) is epistemic, and on this reading the statement clearly licenses free choice inferences, as does the ability modal in (5).
Finally, it has been observed that free choice construals don’t even require
modals (Eckardt 2007, Fox 2007, Klinedinst 2007):
(6)
Many people had pizza or pasta.
; Some people had pizza.
; Some people had pasta.
Surely, (6) doesn’t entail that both pizzas and pastas were consumed, yet this
is what we would normally infer. These observations suggest that the problem
of free choice interpretation doesn’t have anything to do with modality per se,
but is more general.
The free choice inferences observed in the foregoing are of course very similar to the distributive inferences licensed by Boolean “and”, and may therefore
foster the impression that, on a free choice interpretation, “or” is construed as
“and”. This impression is false, however:
6.1 Free choice
(7)
a. You can have fruit or chocolate cake.
b. You can have fruit and chocolate cake.
(8)
a. George may be British or Canadian.
b. George may be British and Canadian.
(9)
a. Many people had pizza or pasta.
b. Many people had pizza and pasta.
107
It is obvious that (9b) is not an adequate way of paraphrasing the free choice
interpretation of (9a), and the same goes for (8a,b). The contrast between
(7a) and (7b) is perhaps less obvious. These sentences are harder to compare
because, even if it is decided that (7a) should have a free choice construal,
both sentences admit of more than one reading. However, (7b) is ill-suited
as a paraphrase for what I take to be the most likely interpretation of (7a),
according to which the addressee can have fruit or chocolate cake but not both.
In sum, free choice phenomena cannot be dispatched by supposing that “and”
occasionally gets mispronounced as “or”.
The problem of free choice inference has prompted a bewildering variety of
responses: logical, semantic, pragmatic, and even syntactic.3 Needless to say,
the course I will be pursuing in this chapter is a pragmatic one. There are at
least three reasons for supposing that free choice inferences are conversational
implicatures. One is that, as a matter of methodological hygiene, we have to
assume that free choice inferences are conversational implicatures, unless there
is compelling evidence that they aren’t. I’ve said it before (§1.3), but the point
is important enough to be reiterated every now and then: if there is a way of
deriving an inference—any inference—by way of conversational implicature,
then it behooves us to do so.
The second reason is that free choice inferences are cancellable:
(10) a. Many people had pizza or pasta—I forget which.
b. George may be British or Canadian, but I’m not allowed to tell you
which it is.
c. You can have fruit or ice cream, but I’m not going to tell you which.
Granted, in some cases the speaker’s motives have to be rather special, if
not downright perverse, for free choice inferences to be cancelled, but as the
examples in (10a-c) show, these inferences are non-conventional, and therefore
a Gricean approach suggests itself.
3.
Recent proposals for dealing with free choice include Zimmermann (2000), Asher and
Bonevac (2005), Geurts (2005), Schulz (2005), Chemla (2007), Eckardt (2007), and Fox
(2007).
108
Chapter 6
The third reason is that some free choice inferences bear a strong resemblance to run-of-the-mill Q-implicatures:
(11) George is British or Canadian.
; He may be British.
; He may be Canadian.
(12) George may be British or Canadian.
; He may be British.
; He may be Canadian.
The only difference between (11) and (12) is that the latter contains a modal
expression which the former lacks; otherwise they are the same. In view of
this similarity, it is natural to expect that (11) and (12) should receive a
unified analysis, and since the inferences in (11) are Q-implicatures, the null
hypothesis should be that the same holds for (12).
6.2 Problems with the Standard Recipe
Oddly enough, while the Standard Recipe explains the inferences in (11) without much ado (as we have seen in §2.3), it dramatically fails to account for the
parallel inferences in (12). To explain this remarkable discrepancy, let’s first
recapitulate how the Standard Recipe deals with (11), as uttered by Mildred
(say):
i. Instead of saying (11), Mildred could have made a stronger statement, like
“George is British” or “George is Canadian”.
ii. Had she believed that either of these propositions is true, Mildred would
have asserted it rather than (11).
iii. Hence, Mildred doesn’t believe that George is British, nor does she believe
that he is Canadian: ¬BelM (George is British) and ¬BelM (George is
Canadian). Or, equivalently, for all Mildred knows, George may not be
British and he may not be Canadian: PossM (¬(George is British)) and
PossM (¬(George is Canadian)).
iv. From (iii ) and the fact that Mildred believes (11), it follows that for
all Mildred knows, George may be British and he may be Canadian:
PossM (George is British) and PossM (George is Canadian).4
4.
If the speaker believes (i ) that ϕ or ψ and (ii ) that ϕ may be false and (iii ) that ψ may be
6.2 Problems with the Standard Recipe
109
The key step in the argument is the first one: it starts with the observation
that Mildred could have made stronger statements than she actually did, on
the basis of which it is inferred that she doesn’t believe that those statements
are true. Formally, this reasoning transforms alternative sentences ϕ into
inferences of the form “The speaker doesn’t believe ϕ”.
The trouble is that when we apply the Standard Recipe to sentences like
(12), the result is nothing short of disastrous:
i. Instead of saying (12), Mildred could have made a stronger statement, like
“George may be British” or “George may be Canadian”.
ii. Had she believed that either of these claims is true, Mildred would have
asserted it rather than (12).
iii. Hence, Mildred doesn’t believe that George may be British and doesn’t
believe that he may be Canadian.
iv. From (iii ) and the fact that Mildred believes (12), it follows that Mildred’s
beliefs about George’s nationality are inconsistent.5
It will be clear that something went seriously wrong, but where? After all, we
employed the exact same reasoning that served us so well with (11), and the
crucial premisses, (i ) and (ii ), wouldn’t seem to be more doubtful than in the
previous example. In fact, the situation is doubly embarrassing: not only does
the Standard Recipe fail quite badly, it also seems out of the question that it
can deliver the unified analysis invited by the obvious similarity between (11)
and (12).
These problems have persuaded some authors that a pragmatic explanation
of free choice is not forthcoming, and that a solution will have to be found
somewhere else altogether (e.g., Asher and Bonevac 2005, Fox 2007). I see no
reason for being so defeatist. In my view, the difficulties caused by free choice
inferences aren’t endemic to the Gricean programme as such, but rather to
the way it is standardly implemented. There is another way, which not only
is more faithful to Grice’s original intentions, but will also give us a unified
analysis of free choice inferences and (other) Q-implicatures.
false, then he has to allow that ϕ may be true (in case ψ turned out to be false) and that
ψ may be true (in case ϕ turned out to be false). More succinctly: if BelS (ϕ ∨ ψ) and
PossS (¬ϕ) and PossS (¬ψ), then it follows that PossS (ϕ) and PossS (ψ).
5. So as to heighten the dramatic impact of the problem, I assume that if Mildred doesn’t
believe that ϕ is possible, then she believes that ¬ϕ. I take it that this assumption is
reasonable, but even if it were dropped the problem would be bad enough.
110
Chapter 6
6.3 Intentions first
When one comes to think of it, there is something odd about the way the Standard Recipe implements the Gricean notion of Q-implicature. The Standard
Recipe begins by asking why the speaker didn’t use this or that linguistic form
instead of the one he did use. But shouldn’t the first question be what the
speaker’s intentional state is, what he wants, believes, etc.? In the following
I will try to show that if we put intentional states first, as any true-blooded
Gricean should, then a unified treatment of free choice and quantity implicatures will fall in our laps, with none of the problems that beset the Standard
Recipe.
I should note that I am using the term “intentional” in the philosophical
sense of the word. Whereas in ordinary parlance this word is related to action, philosophers use it to highlight the directedness of certain mental states:
beliefs, hopes, fears, and desires are all directed at possible states of affairs.
If George believes that it will be sunny tomorrow, the object of his belief is a
proposition, viz. that it will be sunny tomorrow. If Mildred hopes that it will
be sunny tomorrow, her intentional state is different from George’s, but it is
directed at the same proposition. For our purposes, then, intentional states
are mental states directed at propositions.
In the following pages, I will be talking a lot about what speakers might
believe, want, allow, consider possible, as well as about what they might have
said if they were in this or that intentional state. Since it will be vitally
important to separate clearly between intentional states and linguistic means
for expressing them, I would like to emphasise that the noun “alternative” will
be reserved exclusively for referring to linguistic expressions, usually sentences,
that the speaker could have used instead of one he did use.
Scalar implicatures In order to explain the main ideas underlying the intentionbased approach, let’s start with a simple case of scalar implicature. George
says:
(13) Many of the nurses were drunk.
Assuming that George is competent, Mildred infers from his utterance that
BelG (¬(all nurses were drunk)). According to the Standard Recipe, Mildred
achieves this by asking herself why George didn’t make a stronger statement,
i.e. “All nurses were drunk.” She then conjectures that he didn’t do so because
6.3 Intentions first
111
¬BelG (all nurses were drunk), from which she infers, assuming that George
is competent with respect to the proposition that all nurses were drunk, that
BelG (¬(all nurses were drunk)). What is crucial about this story is that
Mildred’s train of reasoning is supposed to start with the observation that
George could have produced a stronger statement and then proceeds to the
inference that he doesn’t believe that claim to be true. In brief, the derivation
of Q-implicatures is driven entirely by alternatives.
To see how it could be otherwise, consider three possible ways George’s
beliefs might relate to the proposition that all the nurses were drunk:
i 1 : BelG (all nurses were drunk)
i 2 : BelG (¬(all nurses were drunk))
i 3 : neither i 1 nor i 2
Each of i 1 -i 3 is a belief state George might be in, and as a matter of logical
necessity, George has to be in one and only one of these states. That is, he
must either believe that all nurses were drunk or believe that not all nurses
were drunk or have no opinion one way or the other. Armed with this threeway partition of belief states, we wind back to the scenario in which George
has just uttered (13), and observe that the following alternative statements
were available to him:
(14) a. All nurses were drunk.
b. Many but not all nurses were drunk.
c. Many and maybe all nurses were drunk.
Given that George wanted to convey, at a minimum, that many of the nurses
were drunk (for that is what he said), he could have uttered (14a) if he was
in belief state i 1 , (14b) if he was in belief state i 2 , and (14c) if he was in
belief state i 3 . As far as I can see, each of these sentences is amongst the
most economical alternatives available to George for conveying explicitly that
he was in the corresponding state (while entailing (13)). For example, if i 1
obtained, George could have said, “Every nurse was sloshed”, but that wouldn’t
have been more efficient than (14a). The same, mutatis mutandis, for (14b)
and (14c).
Now Mildred reasons as follows. Since (14b) or (14c) are both more complex, hence less economical, than (14a), George would need a better excuse
for avoiding the latter than either of the former. If George believed that all
nurses were drunk, there would in general be no good reason for him uttering
112
Chapter 6
(13) rather than (14a). Put otherwise, between the states i 1 , i 2 , and i 3 , the
first one would have been the most likely to give rise to an utterance different
from the one that was in fact produced, and therefore i 1 is the first candidate
for elimination. Once i 1 is discarded, Mildred has decided that George’s belief
state is either i 2 or i 3 , or equivalently, that ¬BelG (all nurses were drunk);
so at this point she has derived a weak scalar implicature. If, furthermore,
Mildred is prepared to suppose that George is competent with respect to the
proposition that all nurses were drunk, i 3 drops out, too, and she is left with
i 2 , which yields the strong implicature that BelG (¬(all nurses were drunk)).
Thus, intention-based reasoning replicates the results obtained previously by
means of the Standard Recipe.
It is important to see exactly how the two methods differ from each other.
The Standard Recipe begins with a set of one or more alternative sentences.
For each of these, the hearer asks himself why the speaker didn’t utter it, and
the answer is that he didn’t because he doesn’t believe that the sentence in
question is true. The intention-based procedure, on the other hand, starts with
a set of possible intentional states, and then proceeds out to eliminate those
that are unlikely to obtain for some reason or other. The intentional states
that survive this weeding process are the ones that, for all the hearer knows,
the speaker might be in. In the example above, there is only one such state,
but note that, without the Competence Assumption, Mildred would have been
left with two possible belief states for George to be in.
In the intention-based analysis, alternatives are less central than in the
Standard Recipe, but they continue to play a very important part. If the hearer
feels that, had the speaker been in state i , he would have spoken differently
than he did, then he will discard i from his roster of possible intentional states.
In brief, alternatives are used for selecting between possible intentional states.
Simple disjunctions Suppose now that we have two atomic propositions: that
George is British and that George is Canadian. Suppose, moreover, that we
don’t know anything about Mildred’s beliefs, as yet, so she may consider it
possible or not possible that George is British. Since the same holds for the
proposition that George is Canadian, we have four possible belief states (assuming that George can have dual citizenship):
i1
i2
i3
i4
: PossM (George
: PossM (George
: ¬PossM (George
: ¬PossM (George
is
is
is
is
British)
British)
British)
British)
∧ PossM (George is
∧ ¬PossM (George is
∧ PossM (George is
∧ ¬PossM (George is
Canadian)
Canadian)
Canadian)
Canadian)
6.3 Intentions first
113
Each of i 1 -i 4 is a belief state Mildred might be in vis-á-vis the proposition that
George is British and the proposition that George is Canadian, and Mildred
has to be in one and only one of these states: i 1 -i 4 are mutually disjoint and
jointly exhaustive.
Now Mildred issues the following statement:
(15) George is British or Canadian.
Assuming that Mildred believes what she says (i.e., that she is not violating
the Quality maxim), i 4 can be discarded rightaway, for if she was in this
state, Mildred would believe that George is neither British nor Canadian. For
the remaining states, the following alternatives would have been available for
Mildred to utter:
(16) a. George is British.
b. George is Canadian.
c. George may be British, he may be Canadian, and he is at least one
of the two.
Given that Mildred wanted to convey, at a minimum, that (15) is true, she
could have uttered (16a) if she was in belief state i 2 , (16b) if in belief state i 3 ,
and something along the lines of (16c) if she was in belief state i 1 . (If the last
sentence seems too prolix, bear in mind that it has to entail (15) as well as
describe i 1 .) It will be clear that expressing i 1 would have been much more of
a hassle than expressing either i 2 or i 3 , and therefore i 2 and i 3 are the prime
candidates for elimination. Once these states are eliminated, only i 1 is left
for Mildred to be in, which is to say that, for all she knows, George may be
American and he may be Canadian.
Summing up: The intention-based method employs the following ingredients:
• A sentence ϕ, as uttered by a speaker S.
• A partitioning i 1 , ..., i n of S’s possible intentional states. Intentional states
may be partitioned in many different ways. In the foregoing examples, it was
the speaker’s beliefs that were at issue, but in other cases different flavours
of intentionality will be relevant (e.g., free choice permission is not about
S’s beliefs; see below).
• Alternatives: sentences S could have used instead of ϕ, had he wanted to
express any of i 1 , ..., i n . Since we can assume that S intended to signal his
commitment to ϕ, each of these alternatives should entail ϕ.
114
Chapter 6
Next hearer H proceeds to weed out those states in i 1 , ..., i n that fail to
meet certain conditions, which broadly speaking fall into two categories. On
the one hand, there are all manner of things H may already know, or at least
assume, about S. For example, if ϕ is a statement, H will assume that BelS (ϕ),
and any of states in i 1 , ..., i n that are inconsistent with this assumption can
be discarded. H may also be assuming that S is competent with respect to
a given proposition, which may cause further casualties amongst i 1 , ..., i n .
More generally, anything H assumes about S’s beliefs, hopes, desires, etc. may
help to narrow down the space of possible intentional states.
On the other hand, H can try to weed out candidate intentional states by
considering the alternatives associated with them, using the following alternative heuristics:6
Alternative heuristics
i. If S could have expressed that he is in state i by using a sentence that is
no more complex than ϕ, then probably S is not in state i .
ii. For any pair of intentional states, i and j, if it had been easier for S to
express that he is in i than that he is in j, then S is more likely to be in j
than in i.
Again, it is presupposed that S’s utterance of ϕ is sincere: if ϕ is a statement,
then S believes ϕ; if ϕ is a permission sentence, S intends to grant H the
permission that ϕ; and so on. Hence, any alternative must be at least as
strong as ϕ.
Both alternative heuristics are based on the idea that there is an inverse
correlation between an alternative’s complexity and the likelihood that the
speaker is in the corresponding intentional state. If ψ is among the simplest
alternatives S could have used for expressing that he is in state i , then the more
complex ψ, the more likely it becomes that ψ’s complexity is what explains
why S failed to utter it, thus making it proportionally less likely that i holds.
This idea has a long pedigree in the history of Q-implicatures,7 but it raises
some hairy issues, which will be addressed in §6.5.
6.
This is to be understood as a compound nominal, i.e., it is “alternative heuristic”, not
“alternative heuristic”.
7. See, for example, Strawson (1952), Gazdar (1979), Levinson (1983), Horn (1984, 1989), and
Blutner (2000).
6.4 Free choice explained
115
6.4 Free choice explained
Epistemic free choice One of the signal virtues of the intention-based analysis
of Q-implicatures is that it extends to free choice inferences without further
ado. To explain how, I will begin by showing how the analysis of the simple
disjunction in (15) carries over to its free choice cousin. This time around,
Mildred’s utterance is slightly weaker than previously:
(17) George may be British or Canadian.
As in the case of (15), we begin by observing that Mildred must be in one of
the following belief states:8
i1
i2
i3
i4
: PossM (George
: PossM (George
: ¬PossM (George
: ¬PossM (George
is
is
is
is
British)
British)
British)
British)
∧ PossM (George is
∧ ¬PossM (George is
∧ PossM (George is
∧ ¬PossM (George is
Canadian)
Canadian)
Canadian)
Canadian)
Of these states, i 4 can be discounted straightaway, because it is inconsistent
with Mildred’s believing (17). For the remaining states, the following alternatives would have been available for Mildred to utter:
(18) a. George may be British.
b. George may be Canadian.
c. George may be British and he may be Canadian.
Assuming that Mildred believes what she says, she could have uttered (18a)
if she was in belief state i 2 , (18b) if in belief state i 3 , and something along
the lines of (18c) if she was in belief state i 1 . Now the second alternative
heuristic tells us that, since (18a) and (18b) are considerably simpler than
(18c), Mildred’s intentional state probably is i 1 rather than i 2 or i 3 , which is
to say that, as far as Mildred knows, George may be British and George may
be Canadian. Free choice explained.
Thus the intention-based theory of Q-implicature explains how (17) gives
rise to free choice inferences and, moreover, the explanation it offers is prac8.
I’m assuming here that the propositions at issue are that George is British and that he is
Canadian, not that he may be British and that he may be Canadian. I believe that this is
a natural assumption to make, but even if it simplifies things somewhat, it is not critical,
as will become clear presently.
116
Chapter 6
tically the same as the one it gives for simple disjunctions. It is a unified
treatment if there ever was one.
Let’s try to pinpoint why the intention-based account succeeds where the
Standard Recipe failed. Consider the alternative in (18a). In a sense, the two
theories agree that this alternative is to be rejected, but they disagree on what
exactly this means. According the Standard Recipe, what it means is that
the hearer is entitled to infer that Mildred doesn’t believe that (18a) is true—
which is wrong. On the intention-based account, by contrast, what is rejected
is a possible belief state that would have caused Mildred to utter (18a), to wit:
that she believes that George is not Canadian but may be British—and this
prediction is correct. Herein lies the main difference between the two accounts.
Once we are clear about the distinction between alternatives and intentional states, it is obvious enough where the Standard Recipe went wrong:
it conflates alternatives with the objects of possible intentional states. It
is this conflation that obstructed a unified treatment of free choice and Qimplicatures.
Free choice permission Now let us return to the opening example of §6.1: my
son asks, “What’s for dessert?”, and I say:
(19) You can have fruit or chocolate cake.
There are various ways of construing this answer, but let’s suppose that the
modal is being used deontically and performatively: I’m granting my son permission to have fruit or chocolate cake. Thus construed, the deontic case has
the same logical structure as the epistemic one. True, (19) speaks of desserts
rather than citizenship and its modality is deontic instead of epistemic, but
none of this makes any difference for the computation of free choice inferences.
Suppose that, as far as my son is concerned, I am the dessert authority,
and until I answered his question he had no idea what would be my position
in the matter. Hence, as far as my son was aware, he might or might not be
allowed to have fruit, and he might or might not be allowed to have chocolate
cake:
i1
i2
i3
i4
: PermB (F)
: PermB (F)
: ¬PermB (F)
: ¬PermB (F)
∧ PermB (C)
∧ ¬PermB (C)
∧ PermB (C)
∧ ¬PermB (C)
6.4 Free choice explained
117
where F is the proposition that my son has fruit and C is the proposition
that he has chocolate cake. From this point on, the reasoning is the same
as in the previous example, so we can go through it fairly quickly. (19) is
inconsistent with i 4 , and i 2 and i 3 can be discarded, too, because if either of
them had obtained, I would have said something else, and shorter, than what
I did say, e.g., “You can have fruit” for i 2 and “You can have chocolate cake”
for i 3 . Therefore, my intentional state can only be i 1 , and my son is entitled to
conclude that he can have fruit and that he can have chocolate cake (though
perhaps not both).
Free choice reports Having shown how (19) can give rise to free choice inferences when construed as a permission-giving speech act, let’s consider next
how such inferences can arise when the same type of sentence is used reportatively, as in the following scenario. We’re having dinner in a restaurant, and
my son again wants to know what the dessert options are. I consult the menu,
and then utter (19). This time around, I’m not granting my son permission
to choose between fruit and chocolate cake; rather, what I’m conveying is my
belief that the restaurant permits him to make that choice. Hence, in this
case, we have not one but two layers of intentionality to consider: my beliefs
and the restaurant’s permissions.
So, my son will want to consider my beliefs about two propositions: that
the restaurant permits him to have fruit and that it permits him to have
chocolate cake (of course, his main interest is in the latter proposition, but for
argument’s sake we will pretend that he has at least an academic interest in the
former, as well). If we notate these propositions as PermR (F) and PermR (C),
respectively, my possible belief states can be partitioned as follows:
i1
i2
i3
i4
: PossB (PermR (F))
: PossB (PermR (F))
: ¬PossB (PermR (F))
: ¬PossB (PermR (F))
∧ PossB (PermR (C))
∧ ¬PossB (PermR (C))
∧ PossB (PermR (C))
∧ ¬PossB (PermR (C))
Note that, save for the fact that the objects of Poss are more complex, this is
essentially the same as the partitioning we employed in the case of epistemic
free choice, so my son’s reasoning will be the same, too, and its outcome will
be that, for all my son knows, I must be in belief state i 1 : I consider it possible
that he can have fruit and I consider it possible that he can have chocolate
cake. This is correct as far as it goes, but for my son’s purposes it doesn’t
118
Chapter 6
go far enough, yet, because what he would like to have is that I believe he
can have fruit and (more importantly) that I believe he can have chocolate
cake. Fortunately, he can infer this if he is prepared to assume that I am
competent with respect to the propositions PermR (F) and PermR (C). Since
it is quite plausible that the Competence Assumption holds in these cases (I
did see the menu, after all), it follows that BelB (PermR (F)) and (finally!)
BelB (PermR (C)).
Pizza or pasta To conclude our tour of the free choice garden, let’s see how
the intention-based account applies to examples involving ordinary quantifiers,
like:
(20) Many people had pizza or pasta.
; Some people had pizza.
; Some people had pasta.
Unlike many other accounts, which are often geared to the modal cases (like
Geurts 2005, for example), the present analysis handles this type of example
as a garden variety of free choice, which itself comes out as a garden variety of
Q-implicature. We start with the following partitioning of speaker S’s belief
states:
i1
i2
i3
i4
: PossS (some
: PossS (some
: ¬PossS (some
: ¬PossS (some
people
people
people
people
had
had
had
had
pizza)
pizza)
pizza)
pizza)
∧ PossS (some people
∧ ¬PossS (some people
∧ PossS (some people
∧ ¬PossS (some people
had
had
had
had
pasta)
pasta)
pasta)
pasta)
Applying the same reasoning as in the preceding examples, we draw the conclusion that, most likely, S is in belief state i 1 , which gives us two weak implicatures: that S considers it possible that some people had pizza and that
S considers it possible that some people had pasta. If the Competence Assumption holds with respect to these propositions, it follows from this that
BelS (some people had pizza) and BelS (some people had pasta). These predictions seem to be correct.
Taking stock Prima facie, the intention-based method for deriving Q-implicatures may not be very different from the standard one. Yet, while the key
ingredients—alternatives and intentional states—are the same in both cases,
the ways they are deployed aren’t. The Standard Recipe assumes a close
6.5 Comparing alternatives
119
coupling between alternatives and intentional states: if the speaker didn’t use
alternative ϕ it is always because he doesn’t believe that ϕ. On the intentionbased approach this isn’t necessarily so. For instance, Mildred’s failing to say
“George may be British” may cause her audience to reject the possibility that,
according to Mildred, George may be British but not Canadian. On this view,
the connections between alternatives and intentional states aren’t always as
simple as assumed by the Standard Recipe.
On the intention-based view, the derivation of Q-implicatures is not driven
by alternatives, as it is according the Standard Recipe. However, to say that alternatives are not in the driver’s seat anymore is not to deny their importance;
on the contrary, they are as important as ever, because candidate intentional
states are vetted on the basis of, inter alia, the relative complexity of the alternatives that could have made them explicit. This part of the theory carries
a considerable share of its explanatory burden, and it raises some challenging
issues, as we are about to see.
6.5 Comparing alternatives
To start explaining the issues just alluded to, let’s revisit the analysis of scalar
implicatures of §6.3, where we discussed the following example:
(21) Many of the nurses were drunk. (= (13))
Starting out from a three-way partition of the speaker’s (i.e., George’s) belief
states,
i 1 : BelG (all nurses were drunk)
i 2 : BelG (¬(all nurses were drunk))
i 3 : neither i 1 nor i 2
I argued that i 1 is the prime candidate for elimination, because i 1 could have
been expressed by a sentence (viz. “All nurses were drunk”) that is no more
complex than (21) and less complex than any of the sentences that could have
been used for expressing i 2 or i 3 . If this is right, (a) economy of expression
is a key factor in the derivation of Q-implicatures and (b) when assessing an
intentional state i , the hearer not only has to compare i ’s alternative(s) with
the sentence uttered by the speaker, but may also have to take into account
other possible intentional states and their alternatives. Hence, in general,
candidate intentional states cannot be evaluated individually; rather, they
and their alternatives have to be weighed against each other.
120
Chapter 6
To return to example (21): once i 1 is discarded we are left with i 2 and i 3 .
The alternatives associated with these belief states are about equally complex:
(22) a. Many though not all of the nurses were drunk.
b. Many and perhaps all of the nurses were drunk.
(i 2 )
(i 3 )
Therefore, there is no reason for preferring either belief state to the other on
the grounds that would have been less economical to express. Nevertheless,
I believe that, ceteris paribus, there may be a preference for eliminating i 3 ,
resulting in the strong implicature that, according to George, not all the nurses
were drunk. If there is such a preference, it could be due to a tendency to
assume by default that the speaker is competent (cf. §2.3, §5.4), but this is as
it may be.
The limits of economy Now consider the following:
(23) It must have been a great party. I hear that well before midnight all the
doctors, many though not all of the interns, and (21) many of the nurses
were drunk.
I think that in this case we would be more likely to infer that George doesn’t
know whether all the nurses were drunk, which puts him in belief state i 3 . If
this is so, the reason must be that the discourse in (23) raises the availability of
the “many though not all” construction, which is employed in the i 2 -alternative
in (22a). Therefore, the hearer may reason that if George was in state i 2 , then
given that he used “many though not all” just an instance ago, he should have
uttered (22a) rather than (21). If these observations are on the right track,
what they show is that comparing between alternatives isn’t just a matter of
economy (cf. Katzir 2007). It would have been convenient if we could have
made do with a complexity ordering that ranked alternatives by, say, the
number of syllables they contain, but the example in (23) suggests that this
will not always work.
It is sometimes suggested that economy of expression isn’t implicated in
the derivation of Q-implicatures, to begin with, but it seems to me that this
conclusion is way too strong. While examples like (23) suggest rather forcefully
that economy is not the only factor, they don’t even begin to show that it is
not a factor at all. To the best of my knowledge, the only attempt at arguing
for the stronger conclusion has been made by Matsumoto (1995), using the
following example (p. 44):
6.5 Comparing alternatives
121
(24) It was warm yesterday, and it is a little bit more than warm today.
; It was not “a little bit more than warm” yesterday.
Apparently, the first conjunct of (24) gives rise to a scalar implicature despite
the fact that the relevant alternative (i.e., “It was a little bit more than warm
yesterday”) is clearly more complex. Therefore, according to Matsumoto,
[...] the relative prolixity of [the stronger alternative] cannot by itself constrain the production of [a quantity implicature]. If a stronger item is regarded as carrying necessary information, that expression is expected to be
used even if it is prolix. (ibid.)
Matsumoto’s second claim is clearly correct: necessary information must be
given no matter how much effort it will take. But the first claim is a dubious one. Observe, first, that the preferred interpretation of (24) is forced
by juxtaposing “warm” with “a little bit more than warm”, which raises the
possibility that it isn’t due to a Q-implicature in the first place. Rather, the
upper-bounding effect may have the same cause as in sentences like (25), which
are generally agreed not be amenable to an analysis in terms of implicature
(see §7.1 and §8.5 for further discussion):
(25) Actually, it wasn’t warm yesterday: it was a little bit more than warm.
Secondly, even if (24) is a case of implicature, it can be explained along the
same lines as (23). On this analysis, it is the speaker’s own use of the expression
“a little bit more than warm” that renders it a viable alternative to “warm”.
If the speaker goes out of his way to use this form in the second half of his
statement, the reason why he did not do so in the first half cannot be that it
was too complex, and thus the scalar inference is accounted for.
Availability I would like to replace economy with more general measure for
comparing alternatives, which I propose to dub “availability” (cf. Horn 2000,
Katzir 2007). Economy is one of the factors that determine availability, but
there are further factors, as well, including previous use, frequency of use, familiarity, and general or particular expectations of specificity. In the foregoing,
we have seen how previous use can enhance the availability of alternatives. I
will give a few more examples. Consider the term “megalopolis”, which means
something like “oversized metropolis”. Given that the word is relatively rare,
and much rarer than “metropolis”, the use of the latter would not normally
implicate that the predicate “megalopolis” doesn’t apply. Hence, if someone
should claim that there is a metropolis on the back of the moon, his statement
would not normally give rise to the implicature that the said city is not a
122
Chapter 6
megalopolis. However, if the word “megalopolis” came to be used more frequently, and thus became more available, such an implicature would be more
likely to occur.
A somewhat more pedestrian case is discussed by Horn (2000), who observes the following pattern:
(26) a.
I hurt my finger.
; I didn’t hurt my thumb.
b. I hurt my toe.
6; I didn’t hurt my big toe.
Although thumbs and big toes are structural analogues, and thumbs surely
count as fingers (“or we’d all have eight fingers instead of ten”; Horn 2000:
308), (26a) is easily read as implicating that the finger in question wasn’t the
speaker’s thumb, while a parallel inference seems rather unlikely in the case
of (26b). Evidently, this contrast is related to the fact that English has a
monolexical expression for thumbs but not for big toes, and Horn conjectures
that, “if the colloquial language replaced its thumb with the polymorphous
pollex (the Latin and scientific English term for both “thumb” and “big toe”),
the asymmetry [between (26a) and (26b)] would instantly vanish.” (ibid.) It is
important to note, however, that the adjective “colloquial” is doing real work in
this statement: it is not enough for an alternative word to be in the language;
it has to be sufficiently salient, as well: if the word “thumb” was rarely used,
then presumably the asymmetry between (26a) and (26b) would vanish, too.
In §2.4, we discussed the interpretation of sibling terms in Japanese (Matsumoto 1995). Recall that, in addition to “kyoodai”, which is synonymous with
English “brother”, Japanese also features words meaning “younger brother”
(“otooto”) and “older brother” (“ani”), and that the use of “kyoodai” will normally be taken to implicate that the speaker doesn’t know which of the hyponyms is applicable. By contrast, the use of “dog” in English does not normally implicate that the speaker doesn’t know which of its hyponyms apply in
the case at hand, even though colloquial English has words like “poodle” and
“terrier”, which are by no means rare. In brief, it seems that Japanese “ani”
and “otooto” are more readily available than English “poodle” or “terrier”, the
reason being, presumably, that to speakers of Japanese seniority is more important than are the differences between dog breeds in most of the Western
world.
What I’m suggesting, then, is that the availability of “ani” and “otooto”
is relatively high because, in Japanese culture, seniority is of general interest,
6.6 Two flavours of Q-implicature
123
and therefore relevant by default. But, again, it doesn’t follow that availability is the same thing as relevance, because it is crucial that Japanese makes a
lexical distinction between older and younger brothers. To see that this matters, compare Japanese “kyoodai” to English “cousin”. Surely, the dichotomy
between men and women is of universal interest (and more important, presumably, than the distinction between older and younger brothers), and yet,
if I say that George will be visiting a cousin of his, there will not normally
be an implicature to the effect that I don’t know whether the cousin is male
or female, though there almost certainly would be such an implicature if the
distinction between male and female cousins was lexicalised in English, as it
is many other languages.
To sum up, I have argued that in the process of weeding out possible
intentional states, the hearer may have to take into account the availability
of alternatives associated with those states, and that availability is not just a
matter of economy, but is also affected by a variety of factors including previous
use, frequency of use, familiarity, and general or particular expectations of
specificity. How precisely these factors conspire to determine the availability
level of an alternative is an issue that will be difficult to resolve, if only because
the various factors just mentioned are interdependent to some degree at least,
and I will not try to sort out these matters here. However, there is a related
point that is at least as important, which is that availability cannot always
play a role in the derivation of Q-implicatures, because in many cases there
are no alternatives to be compared, in the first place. It is to this issue that
we now turn.
6.6 Two flavours of Q-implicature
In the introduction to this chapter we briefly discussed the following example:
(27)
George: What did you have for lunch?
Mildred: I had some of the strawberries. (= (1))
We observed that, aside from scalar “not α” inferences (where α is a quantifier
that outranks “some”), Mildred’s answer might also implicate that all she had
for lunch was some strawberries, and that it seems unlikely that the latter
inference can be accounted for by the Standard Recipe, because this would
involve rejecting an indefinite number of alternatives:
124
Chapter 6
I had some of the strawberries and a biscuit.
I had some of the strawberries and a banana.
I had some of the strawberries and some muesli.
etc.
Apparently, George must be able to infer that Mildred only had strawberries
without having a precise notion of what else she might have had for lunch.
True, George may have good reasons for believing that, e.g., muesli would
have been a highly likely lunch option for Mildred to settle upon, and that
she would never even consider consuming boeuf bourguignon in the middle of
the day. But normally speaking the set of possible lunches is open-ended, and
therefore the same must hold for the set of alternative answers Mildred could
have given.
We might say that Mildred’s answer to George is associated with two sets
of alternatives. First, there is the alternative set involved in the scalar implicatures connected with “some”; this set is more or less clearly defined and fairly
small. Secondly, there is the set of alternatives associated with the “strawberries and nothing else” implicature, which is open-ended and potentially huge.
Let us say that the first alternative set is “closed”, while the second one is
“open”.
The distinction between open and closed alternative sets correlates with
another one. On the one hand, the members of a closed alternative set are,
as a rule, no more complex than the sentence uttered by the speaker. On the
other hand, the members of an open alternative set are typically, though not
perhaps necessarily, more complex than the speaker’s sentence. If Mildred had
wanted to convey that, besides the strawberries, she had a wedge of cheddar,
as well, her answer would have had to be longer, obviously.
Although the correlation between the open/closed distinction and the complexity of alternatives is by no means a logical necessity, there aren’t that
many exceptions I can think of. The examples in (23) and (24) show that the
members of a closed alternative set are sometimes longer than the speaker’s
sentence, but these are out-of-the-way cases, and I can’t think of any good
examples of open alternative sets whose members aren’t more complex than
the speaker’s sentence.9 Be this as it may, it seems quite safe to say that the
correlation is reliable enough.
9.
If the context imposed clear limits on the what Mildred might have had for lunch, she could
have said: “I ate everything.” However, in such a case Mildred’s set of alternatives would
be closed, so this is an apparent exception only.
6.6 Two flavours of Q-implicature
125
Most of the discussion in this book has been about Q-implicatures involving closed sets of alternatives: scalar implicatures, ignorance inferences
associated with disjunction, and free choice inferences. In these cases, there
are always small and determinate alternative sets, and it makes sense to assume that hearers compare alternatives, using the heuristics proposed in §6.3.
But clearly, this presupposes that alternative sets are closed; if a set of alternatives is open, and it is indeterminate what the speaker’s alternatives are,
then the alternative heuristics don’t apply. This conclusion is in line with our
discussion of (27). For, in this case, George simply cannot derive the relevant
Q-implicature by way of rejecting alternatives. Rather, we observed, what he
has to do is ask himself:
(28) If Mildred believed (knew) that her lunch comprised more than just some
strawberries, would she have said so?
And this question he may have to answer without having any idea what Mildred’s alternatives were.
Two flavours of Q-implicature If these observations are on the right track,
there are two rather different flavours of Q-implicature, which in the remainder of this chapter I will distinguish by means of subscripts. Qc -implicatures
are derived on the basis of closed sets of alternatives; they comprise scalar
implicatures, ignorance inferences licensed by “or”, and free choice inferences.
Qo -implicatures, by contrast, do not involve closed set of alternatives. One
specimen of this class is the “strawberries only” inference prompted by Mildred’s answer in (27); further examples will be discussed shortly.
What is the point of distinguishing between Qc - and Qo -implicatures? One
good reason for introducing this distinction is that these two kinds of Qimplicature have somewhat different etiologies: while alternatives play a key
role in the derivation of Qc -implicatures, Qo -implicatures are not contingent
on the availability of alternatives. Basically a Qo -implicature is derived on the
assumption that if there was more to say, the speaker would have done so, no
matter how; unlike Qc -implicatures, Qo -implicatures are not constrained by
the form of what the speaker might have said.
Another reason for distinguishing between Qc - and Qo -implicatures is that
the former appear to be more robust and determinate than the latter. To
illustrate the difference, consider again the free choice inferences prompted by
our mainstay example:
126
Chapter 6
(29) You can have fruit or chocolate cake.
In this case, it is perfectly clear cut what the relevant (Qc -) implicatures are:
you can have fruit and you can have chocolate cake. Qo -implicatures, by
contrast, lack this determinacy, as (27) illustrates. Suppose that, in addition
to a handful of strawberries, Mildred nibbled, say, two or three peanuts or a
diminutive wedge of cheddar. Would that falsify the implicature that she only
had some strawberries? Probably not. If we increase the number of peanuts
or the size of the wedge, we may reach a point at which we start wondering
whether Mildred really had only strawberries for lunch (though it might be
argued that the peanuts don’t count, because they’re not lunch food), but
clearly the cutoff point is nowhere as clear as it is in the case of (29).
To see that (27) is not an isolated case, consider the following examples:
(30) Mildred: What did you do when you got home?
George: I fell asleep in front of the television.
(31)
George: Who did you see at the reception?
Mildred: Jack, Jill, and Janet.
It is rather unlikely that falling asleep in front of the television was the only
thing George did when he got home. He may have taken off his coat and
shoes, made himself a cup of tea, and so on. What George’s answer in (30) is
meant to convey is that falling asleep in front of the television was the only
noteworthy thing he did. But clearly, noteworthiness is a matter of degree,
and therefore the Qo -implicature prompted by George’s answer is much less
determinate than your average Qc -implicature. Similarly, Mildred’s answer
in (31) will not normally implicate that she saw no more than three people.
George is entitled to suppose that, had she seen anybody else of note, she
would have said so, but this is vague in the same way George’s own answer in
(30) is.
The same indeterminacy can be observed in conditional statements like the
following:
(32) If you scratch my back, I’ll scratch yours.
Someone who utters (32) will normally convey that, unless the addressee
scratches her back, she will not scratch his, and I take it that this is a Qimplicature: if there were other conditions under which the speaker would be
willing to scratch the addressee’s back, she would have said so, and since she
didn’t, the addressee can infer that scratching the hearer’s back is a condition
sine qua non. Strictly speaking, however, this can’t be true, because there
6.6 Two flavours of Q-implicature
127
are zillions of other situations in which the speaker would scratch the hearer’s
back:
If you threaten me with a machete, I’ll scratch your back.
If you give me one million euros in cash, I’ll scratch your back.
If I suddenly feel an irrepressible urge to scratch your back, I will do so.
etc.
More likely than not, each of these alternatives is true. Hence, it is more
accurate to say that (32) implicates that, apart from the addressee’s scratching
the speaker’s back, there are no probable circumstances that will cause the
reverse to happen—which, again, is vague.
Qc -implicatures are more determinate than Qo -implicatures because the
latter are driven purely by considerations of relevance, while the former are
constrained, in addition, by the availability of closed sets of alternatives that,
as a rule, are at least as economical as the sentence uttered by the speaker.
Someone who says, “Some of the nurses were drunk”, could have made a
stronger statement without more personal effort and without inflicting greater
processing effort on the hearer, so even if it is doubtful that more information
is needed, considerations of economy (or, more generally, availability) are not
likely to explain why the speaker decided to use “some” instead of a stronger
quantifier. In Qo -implicatures, by contrast, not only is it unclear what the
alternatives are, it also is a practical certainty that whatever additional information the speaker might have wanted to provide would have required more
effort from all parties to the discourse. As a consequence, Qo -implicatures are
driven by considerations of relevance alone, and therefore inherently indeterminate.
Conditional perfection The Q-implicature associated with the conditional statement in (32) is an instance of what, in linguistics, is often called “conditional
perfection”, a term coined by Geis and Zwicky (1971), whose most famous
example is:
(33) If you mow the lawn, I’ll give you $5.
; If you don’t mow the lawn, I won’t give you $5.
In traditional logic, the same pattern of reasoning is known as “denial of the
antecedent”, and is considered to be a logical fallacy. It should be observed,
though, that even if it is not logically valid, this form of reasoning is pragmatically valid in a great many cases, so writing it off as a fallacy is not particularly
helpful.
128
Chapter 6
Although it is widely agreed that conditional perfection (to stick to the
linguistic terminology) is an implicature, there is no consensus on what sort of
implicature it is. The suggestion that it might be a Q-implicature goes back
to Horn (1972) and Noordman (1977) and more recently has been defended
at length by van der Auwera (1997a, 1997b), among many others. This is my
view, as well, but it is not without its detractors. Indeed, it is opposed by
a more recent persona of Horn’s (2000), whose criticism focuses on van der
Auwera’s proposal. I believe that the dispute between van der Auwera and
Horn is based on a misunderstanding, which is that both parties in the debate
fail to appreciate the Qc /Qo -distinction. Let me dwell on this point for a few
paragraphs.
Van der Auwera construes the inference in (33) as a Q-implicature deriving
from the fact that the speaker could have made any number of stronger claims:
– If you mow the lawn or wash the car, I’ll give you $5.
– If you mow the lawn, wash the car, or do the dishes, I’ll give you $5.
– If you mow the lawn, wash the car, do the dishes, or prepare dinner, I’ll give
you $5.
...
Each of these alternatives entails (32), but not vice versa, and if each of these
gives rise to the implicature that the speaker doesn’t believe that it is true,
then it follows that unless the addressee mows the lawn, the speaker will not
give him $5.
Horn (2000) criticises van der Auwera’s proposal on two counts. First, he
notes, as we have done in the case of (27), that the alternatives hypothesised
by van der Auwera don’t seem to figure explicitly in the reasoning underlying
(33); quite simply, it is highly implausible to assume that the addressee has to
go through all these alternatives just to arrive at the conclusion that he will
have to mow the lawn if he wants to earn $5. Secondly, Horn observes that
van der Auwera’s alternatives are “invoked here as dei ex machina to assure a
successful [...] account” of conditional perfection (p. 309); that is to say, van
der Auwera’s story is ad hoc. Having shown to his own satisfaction that van der
Auwera’s proposal is flawed, Horn decides to drop the idea that conditional
perfection is a garden variety of Q-implicature, and proceeds to develop an
alternative account in which several maxims collaborate to secure the coveted
inference. However, I will not discuss Horn’s own proposal here, because in my
opinion he is too quick to dismiss the feasibility of a straightforward Q-based
explanation.
6.7 Conclusion
129
Despite their disagreement over the true nature of conditional perfection,
van der Auwera and Horn share the view that deriving Q-implicatures is a
matter of dismissing alternatives. What they disagree about it just whether
or not conditional perfection can be explained in these terms. In this chapter
I have argued at length that the alternatives-first view is wrong, and that in
many instances of Q-implicature alternatives aren’t involved at all. If this
much is accepted, it is hard to avoid the conclusion that conditional perfection
is a perfectly ordinary case of Q-implicature. Hence, I partly agree with Horn
as well as van der Auwera. I believe the latter is right in claiming that conditional perfection is a Q-implicature, and that the former is right in criticising
the specifics of the latter’s proposal. On the other hand, both parties share
the tacit assumption that if conditional perfection is a Q-implicature, then it
will have to be of the Qc -type. This is not the case: conditional perfection is
a Qo -implicature.
6.7 Conclusion
In this chapter I have argued that the proper treatment of Q-implicatures
should be founded on two key distinctions. The first and most important
distinction is that between the speaker’s intentional states and the alternative
sentences he could have used. It is obvious enough that these are entirely
different things, but the standard view on Q-implicatures comes dangerously
close to conflating the two, and at the very least it assumes that they are tightly
coupled: if the speaker didn’t say ϕ, then it must be because he doesn’t believe
that ϕ.
On the intention-based view, hearers do not primarily reason in terms of
alternatives but rather in terms of intentional states the speaker might be in.
Still, alternatives continue to play an important ancillary role in the theory:
they are used to weed out candidate intentional states. However, the availability of alternatives is not a prerequisite for Q-implicatures to be derived.
Thus, the intention-based approach prompts the introduction of a second distinction, between Qc - and Qo -implicatures, which we couldn’t make as long as
we were adhering to the view that Q-implicatures must be derived according
to the Standard Recipe. This distinction yields the following taxonomy of
Q-implicatures:
130
(34) Qc -implicatures
a. Many of the nurses were drunk.
; Not all the nurses were drunk.
b. You may have fruit or chocolate cake.
; You may have fruit.
; You may have chocolate cake.
Chapter 6
[scalar implicatures]
[free choice inferences]
(35) Qo -implicatures
a. [What did you do last night?]
[Q/A-implicatures]
I fell asleep in front of the television.
; That’s all I did.
b. If you scratch my back, I’ll scratch yours.
[conditional perfection]
; If you don’t scratch my back, I won’t scratch yours.
The main argument in favour of the intention-based view is that it is more
general than the standard, alternatives-first approach. The Standard Recipe
works well for scalar implicatures, but doesn’t provide a plausible account of
Qo -implicatures, and it fails quite miserably with free choice inferences.
Embedded implicatures:
the problems
7
Even before the William James Lectures had started appearing in print, Grice’s
critics were arguing against his theory of conversational implicature using examples in which, prima facie, conversational implicatures occurred within the
scope of conditionals and other operators; which on a Gricean view amounts to
a contradiction in terms, since conversational implicatures can only be derived
on the basis of a full-blown speech act. The first author to raise this issue was
Cohen (1971), who presented cases like the following:
(1)
If the old king has died of a heart attack and a republic has been declared,
then Tom will be quite content.
Cohen observed that (1) may be understood as:
(2)
If the old king has died of a heart attack and afterwards a republic has
been declared, then Tom will be quite content.
If it is assumed that, within the Gricean scheme of things, the sequential reading of “and” is due to a Manner implicature, this interpretation is problematic,
because it would seem to require that the implicature must be derived within
the antecedent of the conditional; which is simply impossible (see Carston
2002 for extensive discussion of the pragmatics of conjunction). Hence, Cohen
concluded, we have to accept that, in sentences like (1), the truth-conditional
meaning of the word “and” has a temporal component.
After some initial skirmishes following in the wake of Cohen’s article (e.g.,
Walker 1975, Wilson 1975), the discussion about “embedded implicatures” petered out, but it became livelier again when Landman (1998), Levinson (2000),
Chierchia (2004), and Fox (2007) began calling for drastic departures from the
Gricean party line, provoking defensive responses from Sauerland (2004), van
Rooij and Schulz (2004), Horn (2006, 2009), Spector (2006), Russell (2006),
132
Chapter 7
and myself (2009), among others. The current stage of the debate concentrates, as do this chapter and the next, on the status of Q-implicatures, and
more particularly, though not exclusively, the status of scalar implicatures.
Theorists contending that embedded implicatures cannot be accommodated by the Gricean programme raise the spectre of what I call “conventionalism”. In a nutshell, conventionalism is the view that Q-implicatures aren’t
implicatures at all, but rather soft entailments projected from the lexicon or
the grammar. In its simplest form, the conventionalist view is that “some”
really means “some but not all”. Of course, the “not all” part of the meaning
of “some” is cancellable, which is what makes the entailment “soft”, but the
leading idea is that it is included in the lexical content of the word nonetheless.
Alternatively, some conventionalists hold that scalar inferences are generated
in syntax, by way of a covert operator that may be freely inserted into the
parse tree. Either way, the upper-bounding inference associated with “some” is
recategorised as a non-pragmatic aspect of meaning whose origin lies in some
linguistic convention or other—whence “conventionalism”.
If conventionalism is right, the upper bound associated with “some” may go
directly into the truth-conditional content of the sentence at the point where
the word occurs; and if the word is sitting in the scope of another expression,
that is where the upper bound will be inserted, too. Hence, on this view,
“embedded Q-implicatures” are in fact lexico-syntactic enrichments that are
factored in locally.
The ensuing discussion of embedded implicatures will extend over two chapters and fall into three main parts. First, I will survey the problem area and
discuss a variety of cases that, prima facie, qualify as embedded implicatures
(§7.1). Next, I turn to conventionalist theories (§7.2) and argue at some length
that they are on the wrong track (§7.3). Finally, in the following chapter, I
try to show that embedded implicatures can be explained on strictly Gricean
principles, after all.
There are two main reasons why this discussion is going to take so long.
One is that conventionalism is not one doctrine but many, so it will take
a while to sort through the various flavours, and to show that all of them
are problematic for some reason or other. The second reason is that, on the
Gricean view to be defended in the following chapter, there isn’t just one but
several problems of embedded implicatures. More concretely, I will argue that
there are various ways in which pragmatic inferences may arise that seem to
be embedded within the scope of, e.g., an existential quantifier, but in reality
aren’t embedded in any sense, and discussing all of these various ways will
7.1 The problems
133
keep us occupied for the best part of the chapter.
Before we get to work, a couple of preliminary remarks are in order. First,
although I have just spent an entire chapter advertising an intention-based
approach to Q-implicatures, to replace the established alternative-based view,
we cannot ignore the latter entirely, for the simple reason that some versions
of conventionalism are alternative-based. For the most part, however, the differences between the two approaches don’t matter for the following discussion,
an important exception being the analysis of existential expressions of §8.4,
which hinges on the intention-based view.
One theme that will reappear in the following is defaultism, the reason
being that there is a natural affinity between conventionalism and defaultism.
However, although at least some varieties of conventionalism squarely run into
the problems discussed in Chapter 5, I propose to pretend that discussion
never happened, partly because not all species of conventionalism are alike in
this respect, and partly because there will be other issues that demand our
attention.
Finally, since on a proper Gricean construal, “embedded implicatures” is a
contradiction in terms,1 my use of this term is not to imply that implicatures
can be embedded. However, since it has gained some currency as a descriptor
of a problem (or rather, a collection of problems), that’s how I will use it, too,
though I will continue to employ scare quotes as a reminder that the term
should not be taken literally.
7.1 The problems
Belief reports A good way of introducing the problems raised by embedded
Q-implicatures is by way of belief reports like the following:
(3)
Tony believes that Cleo had some of the figs.
Let’s say that (3) is uttered by one Julius. Then it may well give rise to the
inference that, according to Julius, Tony believes that Cleo didn’t have all of
the figs. This inference involves two layers of belief, so to speak, and is of the
form:
(4)
1.
BelJ (BelT (¬(Cleo had all the figs)))
As is Levinson’s (2000) notion of “truth-conditionally relevant implicatures”, and for much
the same reasons.
134
Chapter 7
The problem is that, if we treat this as regular case of scalar implicature, the
outcome will be somewhat weaker, namely:
(5)
BelJ (¬BelT (Cleo had all the figs))
(Note that (4) entails (5), but not vice versa.) Here is how this implicature
comes about. Upon hearing (3), the addressee can reason as follows:
i. Could it be the case that, for all Julius knows, Tony believes that Cleo
had all the figs, i.e. BelJ (BelT (Cleo had all the figs))?
ii. Presumably not, for then Julius would have said, “Tony believes that Cleo
had all the figs.” Hence, ¬BelJ (BelT (Cleo had all the figs)).
iii. It seems reasonable to suppose that Julius is competent with respect
to the proposition that BelT (Cleo had all the figs). That is, either
BelJ (BelT (Cleo had all the figs)) or BelJ (¬BelT (Cleo had all the figs)).
iv. If so, it follows that BelJ (¬BelT (Cleo had all the figs)) (= (5)).
This result is not wrong, and may be quite satisfactory in many cases, but it
falls short of the coveted inference in (4).
On reflection, it isn’t surprising that, as things stand, this inference is out of
reach for the Gricean approach. All scalar implicatures we have encountered so
far are either of the form ¬BelS (ϕ) or BelS (¬ϕ), where ϕ is a proposition that
is stronger than the one stated by speaker S. In the case of (3), ϕ is BelT (Cleo
had all the figs), so the strongest implicature we can hope to derive is indeed
(5), where the scope of the negation operator is too wide for capturing the
inference in (4).
Prima facie, it looks as if, on the problematic reading of (3), a scalar
inference is derived within the scope of the verb “believes”. Therefore, it seems
that, in order to account for this reading, a conventionalist solution is called
for: if the “not all” inference is attached to the word “some”, then it can be
factored into the truth-conditional meaning of the embedded sentence, so that
(3) simply means that Tony believes that Cleo had some but not all of the
figs, and the problem doesn’t even arise.
One of the potential drawbacks to the conventionalist solution is that it
is prone to overgenerate. Whatever the details of the conventionalist analysis
of (3) may look like, it will yield analogous predictions for attitude reports in
general, and predict that (6a) and (7a) can be read as (6b) and (7b), respectively. Indeed, all conventionalist theories that are currently on the market
predict that these readings should be preferred:
7.1 The problems
(6)
a. Tony hopes that Cleo had some of the figs.
b. Tony hopes that Cleo had some but not all of the figs.
(7)
a. Tony regrets that Cleo had some of the figs.
b. Tony regrets that Cleo had some but not all of the figs.
135
This doesn’t seem to be right, though: it is doubtful that the (b) sentences
are plausible paraphrases of the (a) sentences. The reason why I’m making
this observation here is not that it undermines conventionalism, though it does
that, too. Rather, what I want to point out is that, interestingly, the problem
caused by belief sentences appears to be a local one: it arises with “believe”
and kindred verbs like “think” and “know”, but doesn’t seem to extend to
attitude verbs in general. If this impression is correct, then it may be a bad
idea to deploy general-purpose machinery for generating scalar inferences in
embedded positions, as the conventionalist does. That is of course what I will
be arguing later, but for now my point is just that “believe” and its kin appear
to be special.
Factives I just said that attitude reports with “know” cause the same sort of
trouble as do belief reports. The following variation on (3) shows that they
do:
(8)
Julius knows that Cleo had some of the figs.
Intuitively, it doesn’t seem too far-fetched to construe (8) as conveying that
Julius knows that Cleo had some but not all of the figs. In this respect the
case is no different from (3). But actually (8) is more problematic than (3),
for it may license a further inference, as well, namely that, according to the
speaker, Cleo didn’t have all the figs. Where did that come from?
Chierchia (2004) claims that examples like (8) lend support to the conventionalist approach, but in order to explain his idea I will have to digress a
little into the topic of presupposition. Factive verbs like “know” and “regret”
are presupposition triggers, which is to say that they serve to convey, inter
alia, that the speaker is taking it for granted that their complement is true. If
I say, “Julius knows that Tony is a debaucher”, I’m presupposing that Tony is
a debaucher; what I’m asserting is merely that Julius knows it. In this respect,
factive verbs are similar to definite descriptions, for example, which likewise
trigger presuppositions. For example, if I say,
(9)
Tony didn’t have breakfast with the queen today.
136
Chapter 7
I presuppose that there is a queen and I assert that Tony didn’t have breakfast
with her. (9) also exemplifies a hallmark of presuppositions: they tend to
be invisible to scope-bearing operators like negation. As a rule, if I utter a
sentence ϕ containing an expression that triggers a presupposition ψ, then
it will be as if I had said “ψ and ϕ”, no matter how deeply embedded the
presuppositional expression may be. Hence, (9) will not normally be construed
as, “It isn’t the case that there is a queen that Tony had breakfast with”,
but rather as, “There is a queen and Tony didn’t have breakfast with her.”
Similarly, despite the fact that the definite description is within the scope of
“suspect”, “may”, and “not”, (10a) would normally be construed as conveying
(10b):
(10) a.
Julius suspects that Tony may not have had breakfast with the queen
today.
b. There is a queen and Julius suspects that Tony may not have had
breakfast with her today.
This “projection behaviour”, as it is generally known, makes sense in view of
the fact that presuppositional information is taken for granted. A speaker
who refers to “the queen” doesn’t state that there is a queen; rather, he signals
that he takes it as given that there is a queen, using the definite article to
do so. As the foregoing sentences illustrate, what an utterance says is, at
least in part, based on what it presupposes, rather than the other way round.
Hence, the broad picture, which is widely accepted in the literature, is that
presuppositions come before sentence meaning, which in its turn is the basis
for calculating conversational implicatures.2
It follows from this that there is no way a bona fide implicature could be
included in a presupposition. But that is what we seem to observe in examples
like (8), where it looks as if the inference that Cleo didn’t have all the figs is
part and parcel of presupposition that Cleo had some of the figs. How could
that be? Chierchia’s answer is straightforward: the scalar inference is not an
implicature. Rather, “not all” is part of the lexical content of “some”; therefore,
“Cleo had some of the figs” already entails that she didn’t have all the figs even
before the sentence combines with the main verb; therefore, the sentence as
2.
See, e.g., Geurts (1999a) or Beaver and Geurts (2010) for discussion and references. To say
that this picture is widely accepted is to implicate, of course, that there is no full consensus.
The most prominent detractor is Grice himself, who held rather idiosyncratic views on
presupposition (Grice 1981).
7.1 The problems
137
a whole presupposes that Cleo had some but not all of the figs. Thus the
conundrum posed by (8) is solved courtesy of a non-Gricean analysis of scalar
inferences.
Disjunction The data discussed so far were problematic for the Gricean theory of Q-implicatures because it yielded predictions that were too weak. The
following examples are potentially more serious, for in these cases the Gricean
approach seems to yield incorrect predictions. To begin with, there is a problem with disjunction (Chierchia 2004, Sauerland 2004):
(11) Cleo either had a mango or some of the figs.
One of the stronger alternatives available to someone who utters (11) is:
(12) Cleo either had a mango or all the figs.
If the Gricean account is right, we should expect that (11) can give rise to
the implicature that, according to the speaker, (12) is false, and if (12) is
false, then Cleo had neither a mango nor all of the figs, which entails that
she didn’t have a mango. But that prediction is patently wrong. Prima facie,
conventionalist theories make much more sensible predictions in this case: if
“some” is construed as “some but not all”, (11) comes out meaning that Cleo
either had a mango or some but not all of the figs. Not only is this a plausible
interpretation of (11), it also fails to predict that (12) is false, which is a good
thing.
Indefinites A second type of case in which Gricean reasoning seems to produce
incorrect results is the following. Suppose Tony wants to regale us with an
anecdote about a particular Dutch sailor, Julius, and he starts off like this:
(13) Last year, a Dutch sailor showed some of the symptoms of delirium.
If the Gricean account is right, we should expect that (13) may give rise to
the implicature that, according to Tony, (14) is false:
(14) Last year, a Dutch sailor showed all the symptoms of delirium.
But if (14) is false, then last year no Dutch sailor showed all the symptoms of
delirium, and this is plainly wrong: Tony’s story is about Julius, and was not
intended to have implications for Julius’s fellow sailors.
As in the disjunctive example in (11), a non-Gricean approach seems to
deliver better predictions in this case. If the scalar inference is folded into the
truth-conditional content of the verb phrase, (13) will end up meaning that,
138
Chapter 7
last year, a Dutch sailor showed some but not all of the symptoms of delirium. This interpretation is a sensible one, but unattainable to an implicature
analysis—or so it would seem.
Contrastive contexts The problem cases to be discussed in the remainder of
this section are very different from the ones mustered so far, in that they involve
quite marginal, and therefore strongly marked, uses of scalar expressions. This
may be the reason why they haven’t played much of a part in the back-andforth between Gricean and conventionalist theories. Still, I believe these data
deserve more attention than they have received, because they provide stronger
prima facie support for the conventionalist approach than anything we have
seen in the foregoing, though in the end I will argue that it doesn’t get these
facts right, either. But that argument will have to wait until the next chapter
(§8.5).
By way of introduction, let us briefly recap from §3.2 some facts about
negation:
(15) a. We like coffee.
b. We don’t like coffee.
On the classical Gricean construal of scalar expressions (Horn 1972), the truthconditional meaning of (15a) doesn’t exclude the possibility that we love coffee,
though it may be implicated that that isn’t the case. On the Gricean account,
(15b) simply means that we aren’t sufficiently fond of coffee to warrant the
claim that we like it, which entails that we don’t love coffee, either. The
diagram in (16) depicts how, according to this analysis, the truth-conditional
meanings of “like coffee”, “love coffee”, and their respective negations relate to
each other:
(16)
◦ . . . . . don’t like . . . . .
like
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
◦
don’t love
love
The negation of “like coffee” denotes all individuals outside the denotation of
that predicate, which is to say that, since the meaning “like coffee” doesn’t
impose an upper bound on its denotation, someone who is in the denotation
of “don’t like coffee” cannot be in the denotation of “love coffee”, either. Now,
consider how the picture would change if “don’t love coffee” was entailed by
the truth-conditional content associated with “like coffee”, as it might be on a
conventionalist analysis:
7.1 The problems
(17)
139
. . . . . . don’t like . . . . . .
◦ . . . . . don’t like . . . . .
like
love
◦ . . . . . . . . . . . . . don’t love . . . . . . . . . . . . .
If this were the case, someone who doesn’t like coffee might well be a coffee lover, and (15b) would not rule out the possibility that we love coffee,
which doesn’t seem to be right. Hence, the standard interpretation of negated
sentences like (15b) buttresses the Gricean view that the meanings of scalar
expressions such as “like” don’t impose an upper bound on their denotations.
Unfortunately, there is a fly in the ointment, as the following examples
collected by Horn (1989: 382) demonstrate:
(18) a. Around here, we don’t like coffee, we love it.
b. I’m not happy he’s gone — I’m elated.
In (18a), “don’t like” does not rule out “love”, apparently, and likewise in (18b),
the negation of “I’m happy” does not rule out that the speaker is more than
happy. It is hard to see how we can avoid concluding that, in (18a), the truthconditional meanings of “like” and “love” exclude each other, which is to say
that in this case the picture in (17) gets things right and (16) is wrong; the
same, mutatis mutandis, for “happy” and “elated”, as these words are employed
in (18b).3
It is a moral certainty that examples like (18a,b) are not instances of Qimplicature, for the simple reason that these construals are weaker than the
standard reading of negation, as illustrated by (15b). (If it isn’t immediately
clear that this is so, note that, as construed in (16), “don’t like” is more
informative than it is according to (17): its denotation covers a smaller area
in the former case than it does in the latter.) By the same token, the contrast
between (15b) and (18a) is grist to the conventionalist mill. It very much looks
as if in (18a) the truth-conditional meaning of “like” is restricted to “like but
not love”, and this is just the sort of thing conventionalist theories are designed
to account for. True, it remains to be explained why this type of construal is
abnormal, and why in the normal cases, exemplified by (15b), “like” does not
have an upper-bounded meaning, but there may be ways of dealing with those
problems, as we will see in the next section.
The examples in (18) are special in at least two ways. One is that they
impose unusually strict interpretations on the scalar predicates they contain.
3.
Alternatively, it might be suggested that the examples in (18) are instances of “metalinguistic negation” (Horn 1985, 1989). For reasons discussed at length (if not ad nauseam) by
McCawley (1991), Carston (1996), and myself (1998a), I repudiate this option.
140
Chapter 7
The other is that these interpretations are flagged by contrastive stress. Apart
from denials such as (18a) and (18b), there are other constructions that exhibit
these features. Take, for instance, comparatives:
(19) a. I’d rather have a warm bath than a hot one.
b. It’s better to get few of the answers right than to get all of them
wrong.
If the lexical meaning of “warm” didn’t exclude “hot”, then the truth-conditional
meaning of (19a) would be that having a bath that is warm-and-possibly-hot
is preferred to having a hot bath. This would make little sense, because normally speaking one term of a comparison may not be entailed by another, as
the following examples show:
(20) a. ?I’d rather live in France than in Paris.
b. ?Vegetables are healthier than peas.
Therefore, it seems more than likely that, in the context of (19a), “warm”
means “warm but not hot”. The same, mutatis mutandis, for “few” in (19b).
Disjunctions are similar to comparatives in that one term of “or” is not
supposed entail the other (Hurford 1974):
(21) ?Cleo is allergic to peas or vegetables.
Nevertheless, we observe sentences like the following:
(22) a. Is a parallelogram sometimes or always a square?
b. Most or all of the appartment must be infested with cockroaches.
Intuitively, the first option considered by (22a) is that a parallelogram may but
need not be a square. Similarly, it would appear that, in the context of (22b),
“most” has an upper-bounded meaning which excludes “all”. Conventionalist
theories can account for these construals; the Gricean theory of Q-implicatures
cannot.
One last pair of examples (from Horn 2006) before we get to the general
point:
(23) a.
If you’re convicted of a felony, you’ll spend at least a year in jail.
And if you’re convicted of murder, you’ll be executed.
b. If it’s warm, we’ll lie out in the sun. But if it’s very warm, we’ll
go inside and sit in front of the air-conditioner.
Given that murder is a felony, (23a) should make little sense, unless the speaker
meant to convey that a murderer has to spend at least one year in jail before
7.1 The problems
141
being executed. Since this probably not what the speaker had in mind, the antecedent of the first conditional will have to receive an upper-bounded meaning
which leaves murder out of account. Ditto for (23b).
It will be fairly evident, I trust, what the moral of the foregoing observations is going to be: there’s something very special about contrastive sentences.
If you make sure that it is properly contrasted with a stronger scalemate,
“some” can be used to mean “some but not all”, “warm” can mean “warm but
not hot”, etc. And by “mean” I mean the real thing: these are not implicatures,
but honest-to-goodness truth-conditional effects, and therefore they cannot be
separated from the truth-conditional content of the sentence. In this respect,
too, there is a stark contrast between “embedded implicatures” in contrastive
sentences and, e.g., belief reports. For instance, if (24a) is understood as implying that Tony believes that Cleo didn’t have all the figs, we can paraphrase
it as (24b):
(24) a. Tony believes that Cleo had some of the figs. (= (3))
b. Tony believes that Cleo had some of the figs and he believes that
she didn’t have them all.
By contrast, there is no way you could paraphrase (25) along the same lines,
separating “don’t love” from the truth-conditional content of “like”:
(25) Around here, we don’t like coffee, we love it. (= (18a))
Here, the negative statement may be paraphrased as “Either we don’t like
coffee or we don’t love it”, but crucially, this is a disjunctive proposition, whose
second half is not detachable. Put otherwise, on a contrastive construal, a
scalar expression acquires an upper-bounded meaning which is truly local and
enters into the truth-conditional content of the sentence.
Time to take stock. In the foregoing, we have reviewed two types of examples that cause embarrassment to the Gricean theory of Q-implicatures. On
the one hand, there are the contrastive sentences we discussed last. On the
other hand, we have collected a ragbag of non-contrastive cases in which the
Gricean theory either fails to make the right predictions, as in (26) and (27),
or seems to make predictions that are incorrect, as in (28) and (29):
(26) Belief reports:
Tony believes that Cleo had some of the figs.
; Tony believes that Cleo didn’t have all the figs.
(27) Factive verbs:
Tony knows that Cleo had some of the figs.
; Cleo didn’t have all the figs.
142
Chapter 7
(28) Disjunctions:
Cleo either had a mango or some of the figs.
6; Cleo didn’t have a mango.
(29) Indefinites:
Last year, a Dutch sailor showed some of the symptoms of delirium.
6; Last year, no Dutch sailor showed all the symptoms of delirium.
In the next chapter, I will try to show that, initial impressions notwithstanding,
the second batch of observations can be explained in terms of Q-implicature,
after all, but this doesn’t hold for the contrastive cases. Hence, from a Gricean
perspective, there cannot be a unified account of all the data surveyed in this
section. One of the main attractions of the conventionalist approach is that,
perhaps, it can offer such a treatment.
Before we turn to a conspectus of the varieties of conventionalist theories
currently on the market, I would like to enter a caveat. Having gone through a
series of problematic facts, you shouldn’t come away with the impression that
“embedded implicatures” are a rampant phenomenon. They aren’t. For one
thing, contrastive interpretations of scalar expressions are out of the ordinary
by definition: that’s why they are “marked”. For another, each of the examples
in (26)-(29) represents an isolated case. For instance, as pointed out before,
the problem caused by belief sentences like (26) does not extend to attitude
reports with “hope”, “want”, and so on. Therefore, it is still fair to say that on
the whole the Gricean theory of Q-implicatures gets the facts right. All the
putative counterexamples discussed in this section are special in one way or
other. The question is whether conventionalist theories, which were designed
with these special cases in mind, can offer a general account that is anywhere
nearly as convincing as the Gricean one. There are no bonus points for guessing
what my answer is going to be.
7.2 Varieties of conventionalism
The fundamental intuition underlying all conventionalist theories is that scalar
inferences really aren’t inferences, to begin with, and a fortiori they aren’t pragmatic inferences. Rather, a scalar inference is part of the lexico-grammatical
content of a sentence, and not too different from conventional truth-conditional
meaning. Whence the label “conventionalist”.
7.2 Varieties of conventionalism
143
Conventionalist theories come in two main families: lexicalist and syntaxbased. According to theories of the lexicalist persuasion, “some” implies “not
all” by virtue of its lexical content, though of course the implication is cancellable, which regular truth-conditional content is not. Syntax-based theories, on the other hand, maintain that a scalar expression acquires an upperbounded interpretation when it is within the scope of a covert syntactic operator whose meaning is essentially that of overt “only”. When in the scope of
this covert operator, “some” will be interpreted, in effect, as “only some”, and
thus come to exclude “all”.
Lexicalist variants of conventionalism have been around since the early
1970s (Cohen 1971), and since they are relatively easy to explain, we will start
with them. Syntax-based theories are a recent development (Chierchia 2006,
Fox 2007); they are more complex than their lexicalist brethren, but also more
versatile in their predictions, as we will see.
Lexicalist conventionalism According to the beginner’s version of lexicalism,
upper bounds are hard-wired into the lexical entries of scalar expressions like
“many”, “or”, and “like”.4 These upper bounds are cancellable, and therefore
they cannot be part of lexical meaning in the conventional sense, but they are
part of the lexical content of a word. This simple account readily explains
stock-in-trade observations like the following:
(30) Tony had many of the symptoms of delirium.
; He didn’t have all the symptoms.
If one embraces this version of lexicalism, it is natural to assume that strong
implicatures are the norm: if “not all” is pre-packaged with the lexical content
of “many”, it naturally follows that (30) implies by default that, according to
the speaker, Tony didn’t have all the symptoms of delirium. We have seen
4.
What I call “lexicalism” is formally indistinguishable from the view that, e.g., “some” is
ambiguous between a strong and a weak reading, where the former is preferred. The first
explicit defence of the idea that many conversational implicatures are better analysed as
lexical inferences is Cohen (1971, 1977). Cohen didn’t extend this view to scalar implicatures, though he certainly would have done so if prompted. More recently, Levinson (2000)
and Chierchia (2004) have proposed theories which are effectively lexicalist, though for these
authors the mechanism that associates, e.g., “not all” with “some” is not purely lexical. (Incidentally, since his 2004 paper, Chierchia has moved from a lexicalist to a syntax-based
position; see Chierchia 2006, Chierchia et al., to appear). Wainer (2007) presents a formalisation of a lexicalist theory of scalar inferences (among others), using non-monotonic
logic.
144
Chapter 7
in Chapter 5 that there are rather compelling arguments against the notion
that such inferences have default status, but more importantly for the current
discussion, there are problems with embedding, as well:
(31) Tony didn’t have many of the symptoms of delirium.
As discussed in the last section, the preferred reading of (31) is that Tony had
less-than-many symptoms of delirium. The beginner’s version of lexicalism
doesn’t predict this; what it predicts is that the many-but-not-all reading of
(31) should be the default, which is clearly wrong.
There are various ways of getting around this problem, but they all have a
common core: they drop the assumption that the preferred reading of a scalar
expression is determined by its lexical content alone, since examples like (31)
show that the linguistic environment in which such expressions appear has a
role to play, as well. Bearing this in mind, we could amend our initial version
of lexicalism along the following lines:
(32) Enhanced lexicalism (strong version)
A scalar expression α prefers an upper-bounded construal unless this
weakens the interpretation of the sentence in which α occurs.
In (31), an upper-bounded reading of “many” weakens the statement made
by the speaker, and therefore (32) predicts that, in this case, the preferred
construal of “many” is without upper bound. If there is a strong version of
enhanced lexicalism, then of course there must be a weak version, too:
(33) Enhanced lexicalism (weak version)
A scalar expression α prefers an upper-bounded construal provided this
strengthens the interpretation of the sentence in which α occurs.
This is similar to (32) and makes the same predictions in many cases. In particular, both principles suppress upper-bounded readings in downward-entailing
contexts like (31). But they are not equivalent: (33) is the weaker of the two,
since it predicts preferred upper-bounded readings in fewer instances than (32)
does. To illustrate the difference, compare the following sentences:
(34) a.
All the senators supported most of the bills.
?
;
All the senators supported most but not all of the bills.
b. Thirty nine senators supported most of the bills.
?
;
Thirty nine senators supported most but not all of the bills.
Whereas in (34a) “most” occurs in an upward-entailing environment, its context in (34b) is neither upward nor downward entailing. Consequently, a
7.2 Varieties of conventionalism
145
stronger interpretation of “most” will result in a stronger proposition in the
former case, while in the latter case the resulting proposition will be neither
weaker nor stronger: “Thirty nine senators supported most but not all of the
bills” may be true while “Thirty nine senators supported most and possibly all
of the bills” is false, and vice versa. Given these entailment properties, (32)
predicts that an upper-bounded construal of “most” will be preferred in both
cases, whereas (33) only predicts this for (34a). In the next section, I will
argue that both predictions are false, but I’m happy to concede that either
principle constitutes an improvement over the beginner’s version of lexicalist conventionalism. Let us ponder a bit why. The main innovation is that
both varieties of enhanced lexicalism have relinquished the simplistic idea that
scalar inferences are entirely arranged in the lexicon (cf. §3.2). The linguistic
environment in which a scalar expression appears plays an essential part, as
well, which means that the original lexicalist intuition that, e.g., “not all” is
part of the lexical content of “most” will have to be compromised somewhat.
It will have to be compromised more in order to account for the following
examples:
(35) a.
Cleo didn’t have many of the figs.
; She did have some of them.
b. You can have some of the dates.
; You cannot have them all.
None of the versions of lexicalism presented so far can capture these inferences.
The problem with (35a) is that the scalar inference requires that we take into
account the weaker scalemates of “many”: since “many” occurs in a downwardentailing context, (36) would have been a stronger statement,
(36) Cleo didn’t have any of the figs.
and the inference in (35a) results from denying (36). Clearly, any theory which
merely assumes that “many” may or may not have an upper-bounded reading
cannot explain this.
(35b) causes trouble in a rather different way. To see how, consider (37),
which paraphrases the lexicalist interpretation of (35b), with an upper-bounded
construal of “some”:
(37) You can [have some and not all of the dates].
Since the “not all” rider is inserted in the scope of “can” (it has no other place
to go, if it is part of the lexical content of “some”), (37) actually permits the
addressee not to have all the dates, rather than forbidding him to eat the
146
Chapter 7
lot. The heart of the problem is that the preferred reading of (35b) requires
wide-scope negation and narrow-scope “all”: ¬[you can have all]. A lexicalist
analysis obviously cannot provide this.
Syntax-based conventionalism As I hinted already, these problems, but especially the last one, can only be solved by compromising lexicalist principles
even further, which brings us to the syntax-based branch of conventionalism.
Syntax-based theories stipulate that the grammar is equipped with a covert
operator, whose meaning is essentially that of overt “only”. The intuitive motivation underlying this hypothesis is that, as observed by Levinson (1983), at
least some scalar implicatures can be made explicit by means of overt “only”:
(38) a. Julius had some of the symptoms of dengue fever.
b. Julius had only some of the symptoms of dengue fever.
When construed as implying that Julius didn’t have all the symptoms of
dengue fever, (38a) can be paraphrased as (38b), where “only” associates with
“some”. The intuition informing the syntax-driven programme is that, whenever (38a) licenses this scalar inference, its underlying syntactic form resembles
that of (38b) in that (38a), too, contains a form of “only”, albeit a covert one,
which I will render as “ ”.
As it turns out, the interpretation of
is rather tricky (see Fox 2007 for
discussion), but (39) gives the gist of it:
O
O
O
(39) Semantics of
(ϕ) is true iff ϕ is true and every alternative not entailed by ϕ is false.
O
To explain how this works, let’s consider sentence (40a), and assume that the
relevant alternatives, besides (40a) itself, are (40b) and (40c):
(40) a. Many of the barbarians were pacified.
b. Some of the barbarians were pacified.
c. All the barbarians were pacified.
Of these alternatives, the only one that is not entailed by (40a) is (40c), and
therefore, according to the definition in (39), (40a) is true iff (40a) is true
and (40c) is false. Hence, the meaning of (40a) is that many but not all of
the barbarians were pacified.
Although syntax-based versions of conventionalism are considerably more
complex than their lexicalist cousins, they are more powerful, too, simply
because, in general, they will generate more readings for any given sentence,
owing to the fact that
is assumed to be a syntactic operator that may be
O
O
O
7.2 Varieties of conventionalism
147
freely inserted into the parse tree. Hence, there are many different parses for
(41), for example, the most interesting of which are displayed in (42).5
(41) Cleo didn’t have many of the figs. (= (35a))
(42) a. ¬[Cleo had many of the figs]
b.
¬[Cleo had many of the figs]
c. ¬ [Cleo had many of the figs]
O
O
Besides the O-less parse in (42a), there is an analysis in which O takes scope
over the negation and one in which it is the other way round. (42b) expresses
the “many but not all” reading of (41); (42c) represents the marked reading,
which contrary to (42a) and (42b) is consistent with Cleo’s eating all the figs.
Let’s see how these interpretations come about.
In (42b),
combines with ¬[Cleo had many of the figs], so the relevant
alternatives might be the following:
O
(43) a. ¬[Cleo had some of the figs]
b. ¬[Cleo had many of the figs]
c. ¬[Cleo had all the figs]
Since negation reverses the direction of entailment, (43a) is stronger than
(43b), which in its turn is stronger than (43c). Therefore, according to the
semantics of as given in (39), (43b) (= (42b)) is true iff (43b) is true and
(43a) is false, or in other words: Cleo had less than a large proportion of the
figs, but she did have some. This is the unmarked interpretation of (41), but
enriched with a scalar inference that lexicalist theories cannot account for, as
we have seen.
Concerning the syntactic parse in (42c), where combines with [Cleo had
many of the figs], let us suppose that the relevant alternatives are the following:
O
O
O
(44) a. Cleo had some of the figs.
b. Cleo had many of the figs.
c. Cleo had all the figs.
5.
O
In fact, unless -insertion is prevented from applying recursively, every sentence will have
infinitely many syntactic analyses. For example, in addition to the parses in (42), we also
have the following analyses for (41):
OO[Cleo had many of the figs]
O¬O[Cleo had many of the figs]
OO¬[Cleo had many of the figs] ... and so forth
O-recursion tends to be truth-conditionally inert, but it doesn’t have to be. Recursive
application of O is a key ingredient in Fox’s (2007) proposal for dealing with free choice
¬
inferences.
148
Chapter 7
O(44b) is true iff (44b) is true and (44c) is false; that is to say, iff Cleo had
many but not all of the figs. Hence, ¬O(44b) (= (42c)) is true iff Cleo had
either less than a large proportion of the figs or all of them. This is the marked
reading of (41), which would be opportune in the context of (45), for example:
(45) Cleo didn’t have many of the figs: she had them all.
O
In general, -theories will generate a great many readings for any given
sentence. In the case of (41), free -insertion will yield at least three distinct
interpretations, and there is only one scope-bearing expression to reckon with.
Had there been more, the number of readings might have been considerably
higher. In order to convey an impression of how quickly this number will
escalate, consider (46), where “many” occurs in the syntactic scope of three
expressions: “some”, “thought”, and the negation:
O
(46) According to some sources, Tony thought that Cleo didn’t have many of
the figs.
O
If we allow ourselves to use at most two copies of , this sentence will have no
less than sixteen different parses, most of which come with distinct meanings.
Evidently, the formidable predictive power of syntax-based theories comes with
a hefty price tag: most of the readings they predict are not wanted. Hence,
this family of theories is burdened with an equally formidable disambiguation
problem: how to choose amongst all those readings?
There are various ways of dealing with this issue, but for the moment let’s
confine our attention to the only one that has been seriously considered in the
literature. The idea is to adopt what I will call the “Power Principle”:
(47) Power Principle
If we have two readings, ϕ and ψ, and ϕ is stronger than ψ, then ϕ is
the preferred reading.
(Note, incidentally, that both versions of enhanced lexicalism conform to this
principle, too.) In the case of (41), for example, the Power Principle predicts
that the reading in (42b) is preferred to the ones in (42a) and (42c). Surely, this
prediction is at least partly correct: (42c) is a marked reading by any standard,
and should be less preferred than either of the others, which according the
Power Principle it is. In view of the discussion of defaults in Chapter 5, it is
more doubtful that (42b) should be preferred to (42a); we will return to this
issue shortly.
Although the Power Principle takes the sharp edges off the ambiguity problem, it should be noted that it will not solve the problem completely. The
7.2 Varieties of conventionalism
149
trouble is that, in general, amongst the various readings generated by free
-insertion, there will be some that are incommensurable with respect to
strength; or more accurately, there may be pairs of readings that aren’t entailed by a common third reading and neither of which entails the other. In
such cases, the Power Principle will not resolve the ambiguity, and contextual
factors will have to decide which is best.6
To conclude the presentation of the syntax-based approach to “embedded
implicatures”, let us see how it deals with:
O
(48) You can have some of the dates. (= (35b))
Recall that lexicalist versions of conventionalism fail to explain how (48) can
give rise to the inference that the addressee is not allowed to have all the
dates, the reason for this failure being that, if “not all” is associated directly
with “some”, it is bound to remain within the scope of the modal “can”, whereas
on the reading we need to account for, the negation should take scope over
“can”, while leaving the universal quantifier in situ. Which is precisely what
-insertion will give us if it is applied to the entire sentence. To see that
it will, suppose that, apart from (48) itself, there is only one alternative to
consider, namely:
O
(49) You can have all the dates.
O
Then (48) is true iff (48) is true and (49) is false, which is to say that the
addressee can have some of the dates but is not allowed to have them all.
Problem solved.
The common conventionalist core Conventionalist theories of scalar inference
differ from each other along several dimensions, the most important of which
is that some are lexicalist, whereas others are syntax-driven. Either way,
conventionalist theories generate ambiguities which will have to be resolved
somehow; this problem is much worse for syntax-based theories, but it affects
their lexicalist cousins, too. The most promising way of solving, or at least
mitigating, the ambiguity problem is by adopting some version of the Power
Principle, and it turns out that different formulations of this principle may
yield rather different results. The same holds true for the definition of the
-operator (Chierchia et al., to appear). In short, if you go shopping for
O
6.
Actually, the situation is somewhat more complex, because the precise definition of the
Power Principle is a factor, too. But the main point holds true in any case. See Chierchia
et al. (to appear) for discussion.
150
Chapter 7
a conventionalist theory, there are quite a few choices to be made, and the
variety of possible theories may seem a bit daunting.
Still, there are commonalities, too. First off, it bears emphasising that
conventionalist theories are non-pragmatic to the core, and in this sense are
stipulative by nature. Basically, they solve the problem of embedded implicatures by fiat. This is particularly evident in the lexicalist versions of
conventionalism, which stipulate an upper-bounded lexical meaning for each
scalar expression. But it is no less true of syntax-based versions of the conventionalist doctrine. Hence, the conventionalist approach is at a methodological
disadvantage from the start.
The second common thread in conventionalist theories is that, having been
designed specifically for dealing with “embedded implicatures”, they are committed to the view that upper-bounded interpretations systematically occur
in embedded positions. Consequently, by their very nature, conventionalist
theories are prone to overgenerate readings, and typically seek to relieve this
problem by adopting some version of the Power Principle. As a result, these
theories tend to converge on a range of predictions. For example, as we have
seen in the foregoing, most lexicalist and syntax-based theories agree that
denial uses of negation, as in (50), are marked:
(50) Around here, we don’t like coffee, we love it. (= (18a))
That this use of the negation is dispreferred is explained by the fact that, on
this reading, the truth-conditional meaning of “We don’t like coffee” is quite
weak, compared to other possible construals. Hence, the Power Principle may
be called upon to explain why denial readings are exceptional.
Since most conventionalist theories adopt a version of the Power Principle,
they agree that scalar expressions occurring in upward-entailing environments
will prefer upper-bounded construals:
(51) a.
All the ladies were wearing pink gloves or shawls.
; All the ladies were wearing pink gloves or shawls but not both.
b. Most of these children stutter some of the time.
; Most of these children stutter some but not all of the time.
It remains to be seen whether these predictions are correct, but for the moment
my point is merely that this is what most conventionalist theories predict.
Furthermore, many though not all varieties of conventionalism are committed
to the view that the same holds true of non-monotonic environments; that is,
contexts that are neither upward nor downward entailing:
7.3 Against conventionalism
151
(52) a.
Twelve ladies were wearing pink gloves or shawls.
; Twelve ladies were wearing pink gloves or shawls but not both.
b. About half of these children stutter some of the time.
; About half of these children stutter some but not all of the time.
We have seen already that the strong version of enhanced lexicalism predicts
that these inferences hold by default; the weak version does not make this
prediction. As to the syntax-based branch of conventionalism, it depends on
the details of the Power Principle whether or not a theory of this family is
committed to the inferences in (52). On the definition of the Power Principle
given in (47), these inferences don’t follow, but on some other formulations of
the principle they do (Chierchia et al., to appear).
To sum up, while conventionalist theories generally predict that scalar
expressions occurring in upward-entailing environments will have a preference
for upper-bounded construals, some but not all of them make this prediction
for non-downward-entailing environments in general. In the upcoming section,
we will see that both predictions are wrong.
7.3 Against conventionalism
In this section, I will argue that conventionalism is an unpromising doctrine,
even when judged on its own merits. (The rider is not gratuitous, because
it should always be borne in mind that, in the final analysis, conventionalism is not to be judged on its own merits. Rather, the question to ask is
whether conventionalism offers a significant improvement over the Gricean approach to Q-implicatures, for if it doesn’t, we should stick with the latter on
methodological grounds. Cf. §1.3.) I will begin by showing that there are
some phenomena that are beyond the scope of any conventionalist theory, but
are readily explained by implicature theories. This is an important point not
only because it shows that conventionalism has its limits, but also because it
means that conventionalists are compelled to carve up a range of empirical
data for which Gricean theories offer a unified explanation.
Having charted the natural limits of conventionalism, I proceed to show
that conventionalist theories make false predictions in a wide variety of cases,
focusing my attention first on intensional contexts created by attitude verbs
and modals, and then on quantificational contexts. The upshot of this discussion will be that, against the most fundamental conventionalist tenet, upperbounded interpretations do not occur systematically in embedded positions.
152
Chapter 7
While this is bad news for conventionalism, it is good news for Gricean theories, because it is precisely what they predict.
The limits of conventionalism For all its predictive potential, there are a number of phenomena that are beyond the purview of any conventionalist approach. By way of introducing the first such phenomenon, consider the following scenario. Tony asks Cleo, “Which places did you see on your trip to
Italy?”, and she answers:
(53) I saw Naples, Rome, and Ravenna.
Cleo’s statement is likely to license the inference that Naples, Rome, and
Ravenna are the only places she visited. While lexicalist variants of conventionalism will be hard put to account for this inference, it isn’t much of a
problem for syntax-based theories. All they have to assume is that the alternatives to (53) are of the form “I saw x ”, where x is either an Italian city
or a list of Italian cities, and then stick a copy of
in front of the sentence:
with these alternatives, (53) entails that Cleo only saw Naples, Rome, and
Ravenna.7 So far so good. But now consider the following variation on the
example in (53). Tony asks the same question, but now Cleo says:
O
O
(54) Julius and I first went to Naples and Rome together. Then, while he
went to see Milan, I visited Ravenna.
In the given context, this answer can (and probably will) license the same
inference as did the previous one, but now the inference derives from a pair of
sentences, and since the grammar can only insert copies of into the syntactic
analysis of individual sentences, there is no way this inference could accounted
for by standard conventionalist means.
Needless to say (but I’ll say it anyway), this is very odd. Intuitively, it
shouldn’t make a difference for the inference in question whether Cleo chooses
to convey her travel information by way of one sentence or two, and a Gricean
theory need not distinguish between these cases: the inference is derivable as
a regular Q-implicature either way (cf. §1.6). Therefore, whereas conventionalists can claim to have an analysis for the one-sentence case, they will have
to leave the multi-sentence cases to pragmatics; the implication being that the
two cases are unrelated. Which, as I said, is odd.
O
7.
This is merely to say that an analysis along these lines is technically feasible, not that it is
plausible. See the introduction to Chapter 6 and §6.6 for discussion.
7.3 Against conventionalism
153
And that’s not the end of it, as the following stock-in-trade type of example
will serve to show:
(55) Some of the figs are poisoned.
; Not all the figs are poisoned.
Recall how Gricean theories handle this case: first a weak implicature is derived to the effect that ¬BelS (all the figs are poisoned), and if the Competence
Assumption holds, it follows that BelS (¬(all the figs are poisoned)). By contrast, conventionalist theories derive the stronger inference directly, and there
is no way they could derive the weaker one. On a conventionalist construal,
(55) has two readings, “At least some of the figs are poisoned” and “Some
but not all of the figs are poisoned”, and neither of these captures the weak
implicature routinely predicted by the Gricean account.
This is odd, too. Whereas on a Gricean view weak and strong implicatures
are intimately related, for the conventionalist they couldn’t be more different.
On the conventionalist view, strong implicatures aren’t implicatures; rather,
they are aspects of sentence meaning. Conventionalism doesn’t have a story
about weak implicatures, but since they cannot be analysed in conventionalist
terms, it would seem to follow that they have little if anything to do with
strong implicatures. I take it that the Gricean picture is a lot more plausible
than the one painted by the conventionalist.
If weak implicatures cannot be explained by conventionalist means, how
are they to be accounted for? Chierchia et al. (to appear) suggest that they
are to be taken care of by Gricean pragmatics. This suggestion strikes me
as profoundly odd, if only because this impromptu recognition of Gricean
pragmatics would entail massive redundancy. For, if the conventionalist is
willing to leave weak implicatures to pragmatics, they will be forced to concede,
by their own theoretical lights, that whenever a sentence ϕ gives rise to an
inference of the form BelS (¬ψ), where ψ is stronger than ϕ, this inference
can always be explained in two ways: as an entailment or as a Q-implicature.
This corollary is curious enough to be belaboured a bit. Consider sentence (55) again, and suppose it is uttered in a context in which it licenses
the inference that BelS (¬(all the figs are poisoned)). Conventionalist theories propose to explain this in lexico-grammatical terms: on such occasions,
they maintain, (55) means that some but not all of the figs are poisoned,
and it follows, assuming S is sincere, that BelS (¬(all the figs are poisoned)).
However, if we adopt Chierchia et al.’s suggestion, we may as well bypass the
conventionalist machinery, and allow Gricean pragmatics to derive ¬BelS (all
154
Chapter 7
the figs are poisoned), which will be strengthened to BelS (¬(all the figs are
poisoned)) provided the Competence Assumption holds. In short, Chierchia
et al.’s suggestion pulls the rug from under their own theory.
The long and short of it is that conventionalism has its limits: it cannot
aspire to be a wholesale replacement of the Gricean theory of Q-implicature.
Hence, we should ask ourselves which of the frameworks can provide us with
a more coherent treatment of the facts: one that offers an integrated analysis
of weak and strong implicatures and a unified account of one-sentence and
multi-sentence implicatures, or one that does neither?
False predictions: intensional contexts The main problem with the conventionalist approach is that, no matter what version one fancies, it is bound to
generate predictions that are dubious if not downright false. Here are a few
examples:
(56) a.
I hope that some of my relatives will remember my birthday.
?
;
I hope that not all of them will remember it.
b. You have to read some of these papers.
?
;
You’re not allowed to read all of them.
c. I want you to give me some of your biscuits.
?
;
I want you not to give me all of your biscuits.
Conventionalist theories are designed to predict that these inferences hold.
Take (56b), for instance. If we adopt the syntax-based approach, this sentence
will have the following parses in addition to an -less analysis:
O
O
(57) a.
[you have to read some of these papers]
b. you have to [read some of these papers]
O
On the wide-scope analysis in (57a), (56b) rules out the possibility that the
addressee has to read all the papers, which is weaker than the reading associated with (57b), according to which the addressee is obliged to read some but
not all of the papers, and is therefore not allowed to read all of them. Since the
second reading is stronger than the first one, and both are stronger than the
-less reading, syntax-based theories predict that (57b) is the preferred analysis, and hence that the inference in (56b) should hold by default. Lexicalist
theories won’t be able to generate a reading corresponding to the analysis in
(57a), but otherwise they will make the same predictions. In the case of (56a)
and (56c), the story is along the same lines. In short, for all the sentences in
(56), conventionalist theories predict that the preferred reading should be one
O
7.3 Against conventionalism
155
in which “some” is interpreted, in effect, as “some but not all”, and therefore
the mooted inferences should hold by default.
According to my intuitions, these predictions are plain false, and my intuitions are confirmed by experimental data reported by Geurts and Pouscoulous (2009), who found that their subjects endorsed inferences like (56b)
and (56c) less than 20% of the time, on average. Contrary to what conventionalist theories predict, upper-bounded interpretations are strongly dispreferred in these cases. This conclusion is reinforced by a fact we established in
Chapter 5 (§5.3), namely, that the way you probe speakers’ intuitions about
Q-implicatures may have an effect on the outcome. If you ask someone (yourself, for example) whether any sentence containing “some” implies the negation
of the corresponding sentence with “all”, then this very procedure will induce
at least some bias towards a positive response. Taking this fact into account,
20% is likely to be an overestimation, which only makes things worse for conventionalism.
It would seem, though, that there is an obvious way out for conventionalists, since they might argue that, in these particular examples, the default
readings predicted by his theory are cancelled, simply because they are implausible. Unfortunately for the conventionalist, this line of defence will not
wash. The cancellation argument presupposes that preferred scalar inferences
will quietly withdraw whenever they happen to be implausible, and this presupposition is doubtful, as it is contradicted by the behaviour of bona fide
scalar inferences. Let me give a couple of examples to establish this point.
(58) Some of the liberal parliamentarians voted against the bill.
In some democracies, parliamentary fractions tend to vote en bloc. In the
Netherlands, for instance, it would be quite unlikely that some but not all
liberal parliamentarians voted against a bill. Nevertheless, as far as I can tell,
this wouldn’t diminish the likelihood that (58) is interpreted as implying that
the liberal fraction was divided.
(59) In order to prevent the rinderpest from spreading through the herd, some
of the cows were vaccinated.
Since rinderpest is a highly contagious disease, it would be decidedly odd if
only some of the cows were vaccinated, yet that is how we would understand
(59).
(60) Cleo threw all her marbles in the swimming pool. Some of them sank
to the bottom.
156
Chapter 7
No doubt, it would be very odd if some of Cleo’s marbles failed to sink, yet
according to my intuitions that is precisely what (60) conveys.
What these examples show is that genuine scalar inferences are not so
easy to cancel, at all (cf. Geurts 1999a). From a Gricean perspective, this is
to be expected: scalar implicatures arise because the speaker goes out of his
way to make a statement that is weaker than what he could have said with
equal effort, so it stands to reason that it should require special circumstances
to suppress a scalar implicature. But in particular, lack of plausibility will
generally be insufficient for doing so.
Summing up, the cancellation argument is based on the supposition that
the preference for scalar inferences is so weak that it will be overridden whenever the otherwise preferred reading is even mildly improbable. The foregoing
observations show that this supposition cannot be justified on independent
grounds. Therefore, the fact remains that the example sentences in (56) falsify conventionalist predictions.
More false predictions: quantificational contexts According to my intuitions,
none of the following inferences would go through under normal circumstances:
(61) a.
All the farmers in this region own goats or sheep.
6; None of them own both.
b. At least 300 of the farmers in this region own goats or sheep.
6; At least 300 of them don’t own both.
c. Only six farmers in this region own goats or sheep.
6; Only six farmers in this region own goats or sheep but not both.
However, conventionalist theories of scalar inference unanimously predict that
(61a) and (61b) should hold by default, and at least some of them predict this
for (61c), as well. In (61a,b), “or” occurs in an upward-entailing environment,
and therefore all conventionalist theories agree that an exclusive construal
should be preferred. In (61c), “or” occurs in an environment that is neither
upward nor downward entailing, and therefore an exclusive reading is predicted
by, e.g., the stronger version of enhanced lexicalism and by some syntax-based
theories.
These predictions were tested by Geurts and Pouscoulous (2009) in a verification experiment, using sentences like the following:
(62) a. All the squares are connected with some of the circles.
b. There is more than one square that is connected with some of the
circles.
7.3 Against conventionalism
157
All the squares are connected
with some of the circles.
2 true
2 false
Figure 1: Experimental item used in the verification study by Geurts and Pouscoulous (2009).
c.
There are exactly two squares that are connected with some of the
circles.
Note that these sentences are parallel to the ones in (61), and therefore all
conventionalist theories agree that, in (62a) and (62b), there is a preference
for reading “some” as “some but not all”, while at least some brands of conventionalism make the same prediction for (62c). In Geurts and Pouscoulous’s
experiment, each of these sentences was paired with a picture in which the sentence’s classical and conventionalist construals yielded conflicting truth values.
For example, when interpreted according to the conventionalist canon, the sentence in Figure 1, i.e. (62a), fails to match the depicted situation, because one
of the squares is connected to all the circles, but the sentence is true on its
classical reading.
Geurts and Pouscoulous’s findings were as univocal as they get: not a
single one of their participants’ responses agreed with conventionalist predictions. For example, participants invariably gave a positive response in the case
of (62a) (Figure 1), and the same held, mutatis mutandis, for the other conditions, regardless whether they were upward entailing or merely non-downwardentailing. Hence, these results equally disconfirm weak and strong versions of
conventionalism alike.
158
Chapter 7
Thus far, we have been concerned with what we might call “mainstream
conventionalism”: a spectrum of theories which agree that scalar expressions
preferably receive an upper-bounded construal in upward-entailing environments and perhaps in non-downward-entailing environments generally. It
seems that these predictions are wrong. What to do? One option would
be to retreat into a weaker version of conventionalism that refuses to deal
with the ambiguities it begets, and doesn’t make any predictions about which
construal is the preferred one. Leave it to ... pragmatics.
According to this minimal version of conventionalism, each of the sentences
in (62) has several readings on any given occasion, but it is for pragmatics to
decide which is the right one. To the best of my knowledge, nobody has come
forward to advocate such a minimalist take on conventionalism, as yet, and I’m
tempted to say that this is just as well, since this view is too weak to be taken
seriously. Nonetheless, it instructive to consider how minimal conventionalism
might be tested, precisely because it is such a weak position. If this view could
be disproved it would be a crippling blow indeed.
Obviously, in its most austere form, minimal conventionalism fails to make
any predictions at all: the claim that, e.g., (63a) may or may not be read as
(63b) cannot be falsified.
(63) a. All the squares are connected with some of the circles.
b. All the squares are connected with some but not all of the circles.
In order for the theory to have at least a modicum of empirical bite, auxiliary
assumptions will have to be made. One hypothesis that naturally comes to
mind is that native speakers are able to detect the readings that are distinctive
of the conventionalist approach. That is to say, even if speakers don’t prefer
to hear (63a) as synonymous with (63b), it is reasonable to expect them to
realise that the latter sentence expresses a possible construal of the former. If
this auxiliary hypothesis is adopted, minimal conventionalism predicts that,
in a situation that falsifies (63b), native speakers should judge that (63a) is
ambiguous between a true reading and a false one. This prediction was tested
in the last experiment of Geurts and Pouscoulous (2009)
For this experiment, Geurts and Pouscoulous used the same materials as in
the verification study discussed above, but added a third response option: participants were invited to choose between “true”, “false”, and “could be either”.
According to minimal conventionalism, participants should have a preference
for the last option in the upward-entailing conditions, and perhaps in nondownward-entailing conditions generally. In order to assess whether partici-
7.3 Against conventionalism
159
pants can cope with this type of task, in the first place, Geurts and Pouscoulous
included a number of control sentences which were genuinely ambiguous, for
instance:
(64) a.
The circles and the squares are connected with each other.
– The circles are connected with the squares.
– The circles are connected with each other and the squares are connected
with each other.
b. There are green circles and squares.
– There are green circles and there are squares.
– There are green circles and there are green squares.
A sample control item is shown in Figure 2.
The circles and the squares are
connected with each other.
2 true
2 false
2 could be either
Figure 2: Control item used in Geurts and Pouscoulous’s (2009) ambiguity experiment.
The main finding of this study was that, whereas ambiguous controls were
recognised as such in 70% of the cases, the corresponding rate for the nondownward-entailing sentences was 6% while for the upward-entailing sentences
it was 0%. On a further 5% of the non-downward-entailing trials, participants
gave a true/false response that was consistent with a conventionalist construal.
In raw numbers, only 9 out of 88 responses were consistent with minimal
conventionalism. Moreover, all but one of these “conventionalist” responses
160
Chapter 7
were associated with “exactly two”: between them, upward-entailing “all” and
“more than one” prompted a conventionalist response on one trial only.
In short, while they were rather good at picking out undisputedly ambiguous sentences, participants overwhelmingly failed to produce responses consistent with a conventionalist analysis of scalar expressions, either by classifying
target sentences as ambiguous or by assigning them a truth value in line with
a conventionalist construal. In my opinion, this is very strong evidence against
the conventionalist approach to scalar inferences.
7.4 Conclusion
I have argued that the conventionalist approach to scalar inference has two
major empirical issues to contend with. On the one hand, we have seen that
in some cases conventionalist theories fail to make the right predictions. On
the other hand, there are quite a few cases in which conventionalist predictions
are just false. As far as I can see, the first issue is incontrovertible: there is no
way a genuinely conventionalist theory can hope to explain weak implicatures
or multi-sentence implicatures. However, with the second issue we aren’t quite
done yet. We have seen that conventionalist theories make erroneous predictions about scalar expressions that occur in the scope of various quantifiers and
intensional verbs (“hope”, “must”, “want”, and so on). However, our discussion
of these problems proceeded on the tacit assumption that we were concerned
with relatively standard interpretations only. To be sure, this restriction was
entirely reasonable, since the theories under discussion make predictions about
the preferred interpretation of scalar expressions, and the whole point was to
establish that these predictions are wrong.
What complicates the story is that in all the cases in which we don’t normally obtain what conventionalist theories predict to be the default interpretation, this interpretation can be forced, e.g., by employing contrastive stress.
For example, it is possible to construe (65) as meaning that all the squares are
connected with some but not all of the circles:
(65) All the squares are connected with some of the circles. (= (62a))
Similarly, (66) might be used to convey that the speaker wouldn’t like all his
relatives to remember his birthday:
(66) I hope that some of my relatives will remember my birthday. (= (56a))
These observations suggest that, contrary to what we have supposed so far,
7.4 Conclusion
161
conventionalism might be better cast as a theory of marked interpretation.
This suggestion makes sense for other reasons as well. First, as argued in §7.1,
contrastive construals of scalar expressions have to be analysed as contributing
to the truth-conditional meaning of a sentence, and one of the fundamental
tenets of conventionalism is precisely that upper-bounding interpretations have
truth-conditional effects. Secondly, whereas lexicalist versions of conventionalism are most naturally viewed as default theories, syntax-based accounts
would seem to be geared more towards contrastive interpretations. After all,
covert is modeled on overt “only”, which is contrastive if anything is.
However, if conventionalism is to be recast as a theory of marked interpretation, what are we to make of unmarked specimens of “embedded implicature”,
like (67), for example?
O
(67) Tony believes that Cleo had some of the figs. (= (3))
It doesn’t require contrastive stress to interpret (67) as implying that, for
all the speaker knows, Tony believes that Cleo didn’t have all the figs, and
therefore it would seem that cases like this are outside the purview of a theory
that requires upper-bounded construals to be marked. Which brings us back
to a point made at the beginning of the last section: that conventionalism
cannot aspire to offer a wholesale replacement for the Gricean theory of Qimplicatures. This point now applies with a vengeance: if conventionalism is
rebranded as a theory of marked construals (and this is the most promising
strategy for salvaging the theory, as far as I can see), it will ipso facto be of
marginal importance, at the best of times.
To conclude, I have argued that, if it is touted as a general framework
for scalar inference, conventionalism does not acquit itself particularly well.
However, it is still possible that conventionalism may save its skin by scaling
down its ambitions, and reinventing itself as a theory of marked interpretation.
We will return to this possibility in §8.5.
Embedded implicatures:
a Gricean approach
8
In §7.1, we discussed five lines of evidence that, prima facie at least, are at
variance with the Gricean theory of Q-implicature, and more in agreement with
a conventionalist approach. Since then, we have seen that conventionalism is
not without its problems, either, so the obvious thing to do is to have a closer
look at the evidence and re-assess whether a flight into conventionalism is
really called for. In this chapter, I argue that it isn’t, and to explain what the
general form of the argument is going to be, let’s remind ourselves what the
data were. First, there are cases in which the Gricean theory seems unable to
capture inferences that are intuitively real:
(1)
Belief reports:
Tony believes that Cleo had some of the figs.
; Tony believes that Cleo didn’t have all the figs.
(2)
Factive verbs:
Tony knows that Cleo had some of the figs.
; Cleo didn’t have all the figs.
Secondly, there are cases in which the Gricean theory seems to make false
predictions:
(3)
Disjunction:
Cleo either had a mango or some of the figs.
6; Cleo didn’t have a mango.
(4)
Indefinites:
Last year, a Dutch sailor showed some symptoms of delirium.
6; Last year, there were no Dutch sailors who showed all the symptoms
of delirium.
8.1 Disjunction
(5)
163
Contrastive contexts:
Around here, we don’t like coffee, we love it.
6; The extent to which we appreciate coffee is not enough to warrant
the predicate “like”.
In the following, I will argue that these data fall into two classes, which it is
vitally important to keep distinct. The line separating the two runs between
examples (4) and (5); that is to say, while contrastive sentences are in a class
of their own, the others belong together. What unites the latter class is that,
first impressions notwithstanding, they are consistent with the Gricean theory
of Q-implicatures, after all: for each of the examples in (1)-(4), there is a
perfectly adequate explanation in terms of Q-implicature, and I will spend the
better part of this chapter trying to prove so, case by case.
Which leaves us with contrastive sentences. I have already conceded that
examples like (5) are beyond the scope of the theory of Q-implicatures (§7.1),
and granted this is so, it may seem natural to relegate them to some version
of conventionalism. However, for reasons to be explained at the end of this
chapter (§8.5), this is not the course I recommend. Rather, I will argue for an
approach that yields the same readings as conventionalist theories are designed
to deliver, but does so in a truly pragmatic way. I will even argue that, though
the kind of pragmatic reasoning that is needed operates at a different level
from the one at which conversational implicatures are derived, it does have a
natural place in the grand Gricean scheme of things.
8.1 Disjunction
Actually, sentences like (6) pose not one but two challenges to the theory of
Q-implicatures:
(6)
Cleo either had a mango or some of the figs. (= (3))
One is to explain how (say) Julius’s utterance of (6) could give rise to the
implicature that Cleo didn’t eat all the figs. The second is to avoid predicting that, according to Julius, Cleo didn’t have a mango. Here is how these
challenges can be met. When Julius uttered (6), he had at least the following
alternatives at his disposal:
(7)
a. Cleo had a mango.
b. Cleo had some of the figs.
164
Chapter 8
c. Cleo had all the figs.
d. Cleo either had a mango or all the figs.
Each of these sentences entails (6) but not vice versa, and none of them is
more complex. Hence, for each of these sentences, it would be reasonable to
surmise that Julius doesn’t believe it is true:
(8)
a.
b.
c.
d.
¬BelJ (7a)
¬BelJ (7b)
¬BelJ (7c)
¬BelJ (7d)
Now suppose that the Competence Assumption holds for (7c): BelJ (7c) ∨
BelJ (¬(7c)). In conjunction with (8c), this entails that BelJ (¬(7c)), i.e.
Julius believes that Cleo didn’t have all the figs. Thus the first problem is
solved simply by assuming that (7c) is an alternative to (6), which seems
natural enough.
The implicature in (8c) may be strengthened on the basis of the Competence Assumption. What about the remaining implicatures in (8)? Shouldn’t
it be possible that they are strengthened, too? Consider (8a). If Julius were
competent with respect to (7a), i.e. BelJ (7a) ∨ BelJ (¬(7a)), then it would
follow that BelJ (¬(7a)): Julius believes that Cleo didn’t have a mango—
which would be patently wrong. However, and this is crux of the matter,
there are good reasons for doubting that the Competence Assumption holds
in this case.
There are two ways of arguing this point. One is to adopt a line of reasoning
proposed by Sauerland (2004), who argues that the Competence Assumption
cannot hold for (7a) because Julius’s own statement conveys that he is agnostic about the truth (7a). That this is so follows from the fact that Julius
believes that (6) plus the fact that he doesn’t believe that the second disjunct
is true, which is what (8b) says. More generally: given that BelS (ϕ ∨ ψ)
and ¬BelS (ψ) (i.e. PossS (¬ψ)), it must be the case that PossS (ϕ), or equivalently: ¬BelS (¬ϕ). Therefore, on the assumption that Julius’s statement
is truthful, (8b) entails ¬BelJ (¬(7a)), and since we already have ¬BelJ (7a)
(= (8a)), Julius’s utterance implies that ¬BelJ (¬(7a)) ∧ ¬BelJ (7a), which
is the negation of the Competence Assumption for (7a) (see §2.3 for a more
leisurely presentation of this argument).
There is an another and more direct way of showing that the Competence
Assumption cannot hold for (7a), and it goes as follows. If the Competence
Assumption held in this case, then either Julius believed that (7a) is true or
8.2 Belief reports
165
he believed that (7a) is false. However, given that Julius believes that (6),
he should have made a different statement in either case: if Julius took (7a)
to be true, the second disjunct of (6) would have been redundant, and if he
took (7a) to be false, he should have said “Cleo had some of the figs.” Hence,
if Julius is trying to be cooperative, he cannot be competent with respect to
(7a), and it does not follow from ¬BelJ (7a) that BelJ (¬(7a)).
Given that an utterance of (6) implies that Julius is not competent with
respect to the first disjunct, it will be clear that the same holds for the second
disjunct, (7b), and therefore we don’t predict that (8b) will be strengthened,
either. What about (7d), which caused the main problem we set out to solve?
Here the reasoning is essentially the same as for the two disjuncts of (6). The
simplest version of the argument goes like this. If Julius is competent with
respect to (7d), then either he believes that (7d) is true or he believes that
(7d) is false. In the former case, he should have uttered (7d) instead of (6); in
the latter case, which implies that Cleo had neither a mango nor all the figs,
he should have said, e.g., “Cleo had some of the figs.”
On the face of it, it would seem that, if (6) is construed as “Cleo either
had a mango or some but not all of the figs”, “some” is interpreted in situ as
“some but not all”; that is, it looks as if an upper-bounded reading of “some”
is achieved within the scope of “or”. On a Gricean analysis, this is a mirage:
(6) may implicate that, for all the speaker knows, Cleo didn’t have all the figs,
and if it does the speaker’s utterance may be paraphrased as “Cleo either had
a mango or some but not all of the figs”, but the literal meaning of the second
disjunct of (6) is the classical one, and not upper-bounded in any sense.
8.2 Belief reports
What I just said about disjunctions holds for belief reports, as well:
(9)
Tony believes that Cleo had some of the figs. (= (1))
Though it may seem that in (9) “some” can be construed as “some but not all”,
this interpretation, too, may be explained in terms of Q-implicature, without
resorting to the assumption that the embedded sentence, “Cleo had some of
the figs”, has an upper-bounded meaning. One possible line of explanation has
been proposed by van Rooij and Schulz (2004) and Russell (2006). As observed
in §7.1, the strong scalar implicature that standard Gricean reasoning will
deliver for (9a), when uttered by Julius, is that:
166
Chapter 8
(10) BelJ (¬BelT (Cleo had all the figs))
Now, depending on the context, it may be plausible to assume that, for all
Julius knows, Tony knows whether or not Cleo had all the figs:
(11) BelJ (BelT (Cleo had all the figs)) ∨
BelJ (BelT (¬(Cleo had all the figs)))
That is to say, it may be the case that the Competence Assumption holds not
only for the speaker, but for the subject of the belief report, as well. Between
them, (10) and (11) entail:
(12) BelJ (BelT (¬(Cleo had all the figs)))
Thus the seemingly local scalar inference is accounted for on strictly Gricean
principles.1
There is an alternative way of obtaining the same outcome, which is due
to Spector (2006). To explain, consider the following variation on (9):
(13) a. Tony said that Cleo had some of the figs.
b. Cleo had some of the figs.
Intuitively, (13a) may suggest that, when he uttered (13b), Tony believed that
Cleo didn’t have all the figs. Why? Obviously, because it may be assumed
that Tony’s utterance of (13b) implicated that she didn’t. Armed with this
observation, we return to the belief report in (9), and ask ourselves what sort of
evidence Julius may have had for making this claim. In principle, Julius may
have been relying on all kinds of sources, but there is one in particular that
readily comes to mind: he may have known (or at least believed) that (13a) is
true. But we have just seen how this proposition might prompt the inference
that, according to Tony, Cleo didn’t have all the figs, so if an occurrence of
(9) suggests (13a), it may also suggest (12).
Formally speaking, Spector’s analysis is less economical than the one proposed by Russell and van Rooij and Schulz. For the latter to work, we merely
have to assume that the Competence Assumption holds for the subject of the
1.
As pointed out to me by Larry Horn, the reasoning from BelJ (¬BelT (Cleo had all the figs))
to BelJ (BelT (¬(Cleo had all the figs))) is reminiscent of the “neg-raising” interpretation
of sentences like:
(i)
Tony doesn’t believe that Cleo had all the figs.
; Tony believes that Cleo didn’t have al the figs.
It doesn’t follow from this, however, that neg-raising can explain why the implicature is
strengthened, though it may be that implicature strengthening and neg-raising have a common source, i.e. the assumption that the subject of “believe” is competent (cf. Bartsch 1973).
8.2 Belief reports
167
belief report. This may not be a plausible assumption in general, but in the
case at hand it doesn’t seem far-fetched (which is good enough, since the inference we’re trying to capture does not hold by default). Spector’s story is less
straightforward: in order to explain how (9) comes to imply (12), we have to
assume not only that Tony said that Cleo had some of the figs, but also that
his utterance licensed the implicature that she didn’t have all the figs. Still,
it doesn’t seem unlikely to me that, in some cases at least, hearers consider
what may be the evidence on which the speaker bases his statement, and if
the statement happens to be a belief report, Spector’s explanation strikes me
as eminently plausible. Moreover, we don’t have to choose between the two
explanations on offer: since they don’t exclude each other, it is possible that
hearers rely on both lines of reasoning.
It is important to note that, whichever of the two explanations is adopted,
it does not generalise beyond epistemic attitude reports, i.e. statements about
people’s beliefs, knowledge, and so on. As discussed at length in the last
chapter (especially §7.3), conventionalist theories predict that, in principle,
scalar inferences can occur in just about any embedded position, and this
prediction appears to be false: “embedded implicatures” are the exception
rather than the rule. For example, unless forced by contrastive stress on
“some”, the following should not hold:
(14) Tony hopes that Cleo had some of the figs.
6; Tony hopes that Cleo didn’t have all the figs.
In this light, it is a considerable advantage that both Spector’s and the Russell/van Rooij/Schulz proposal are restricted to epistemic reports. To see that
they are thus restricted, consider first what would have to be assumed for making the latter applicable in the case of (14). Supposing, as seems reasonable,
that Julius’s utterance of this sentence gives rise to the implicature that
(15) BelJ (¬HopeT (Cleo had all the figs))
it should be the case that
(16) BelJ (HopeT (Cleo had all the figs)) ∨
BelJ (HopeT (¬(Cleo had all the figs)))
in order for (15) to license the inference that
(17) BelJ (HopeT (¬(Cleo had all the figs)))
However, it is evident that, in general, (16) will not be a particularly plausible
assumption to make; or at the very least, it will be considerably less plausible
168
Chapter 8
than its counterpart in (11).2
Concerning Spector’s proposal, it will not be hard to see that it doesn’t
extend to (14), either. For, if we ask ourselves what evidence Julius might
have had for making this claim, the answer is not going to be that Tony has
said, “Cleo had some of the figs”; so the analogy with the belief report in (9)
breaks down forthwith.
To sum up, as in the case of disjunction, there are ways of explaining how
a sentence like (9) may come to convey that, according to the speaker, Tony
believes that Cleo didn’t have all the figs, and neither explanation requires the
assumption that anything like a scalar inference is derived within the scope of
“believe”. Furthermore, neither explanation carries over to other embedding
constructions, thus safeguarding the Gricean approach against the proliferation
of unwanted readings conventionalist theories are plagued by.
8.3 Factives and other presupposition inducers
One of the issues raised by Chierchia (2004) in his critique of Gricean pragmatics concerns factive verbs:
(18) a. Tony knows that Cleo had some of the figs. (= (2))
b. Cleo had some but not all of the figs.
c. Cleo had (at least) some of the figs.
An utterance of (18a) may convey that, according to the speaker, (18b) holds
as well, and since the sentence presupposes that its embedded clause is true,
it is natural to suppose that the upper-bounding inference is somehow tied up
with the presupposition. That is what Chierchia assumes, too: on his account,
the “not all” inference attaches to the word “some”, and thus is part of the
meaning of the complement clause of “know” even before the presupposition is
triggered. Consequently, the upper bound on “some” projects out along with
the proposition expressed by the factive complement, and (18a) presupposes
(18b) rather than (18c).
While Chierchia’s analysis delivers passable predictions for sentences with
“know” (though see Russell 2006), it doesn’t work for other factive verbs:
2.
In other words, in a context in which (16) holds, (14) should imply (17). As far as I can see,
this possibility is marginal, at best, but in this case my intuitions are rather labile. More
generally, the claim that there is a clear-cut difference between “think/believe/know”, on
the one hand, and the remaining attitude verbs, on the other, may well be too strong, but
even if it is, my main point is not affected.
8.3 Factives and other presupposition inducers
169
(19) Tony regrets that Cleo had some of the figs.
On its most natural reading, (19) means that Tony regrets that (18c), and
conveys, in addition, that (18b). On Chierchia’s analysis, however, the presupposition has to be the same as the object of the attitude verb: Chierchia
predicts that (19) must either presuppose (18b) and assert that Tony regrets
(18b) or presuppose (18c) and assert that Tony regrets (18c). Either way, the
presupposition is the same as the object of Tony’s regret, and therefore the
most likely reading is not accounted for.
The trouble with Chierchia’s analysis is that his conventionalist convictions
lead him to assume that the scalar inferences we observe in (18a) and (19) are
part of the presupposed content from the start. The obvious alternative is to
hypothesise that these inferences derive from what is presupposed. Reverting
to the classical view that presuppositions come before implicatures and that
both sentences presuppose (18c), we apply the standard Gricean reasoning to
this presupposition, and thus arrive at the implicature that, for all the speaker
knows, (18b) holds, as well. In order to explain that (18a) may convey, furthermore, that Tony believes (18b), we can appeal to either of the analyses
discussed in the last section, and since neither analysis carries over from epistemic “know” to emotive “regret”, we can account for the asymmetry between
(18a) and (19) that Chierchia’s proposal is unable to explain.
As we saw in §7.1, the hallmark of presuppositions is that they tend to
escape from practically any kind of embedding. This being so, the hypothesis
I propose is that, whenever a sentence ϕ presupposes that ψ, and ψ may
implicate that χ, then ϕ may carry the same implicature. The following examples confirm this hypothesis:
(20) a. Does Tony know that Cleo had some of the figs?
b. Please make sure that Tony knows that Cleo had some of the figs.
As is usual with a bona fide presupposition, both sentences suggest rather
strongly that Cleo had some of the figs, and they also convey (or, at the very
least, are easily construed as conveying) that she didn’t have all the figs. This
pattern extends to other presupposition triggers, like it-clefts, for example:
(21) a. It was Julius who stole some of Cleo’s shoes.
b. It wasn’t Julius who stole some of Cleo’s shoes.
c. Wasn’t it Julius who stole some of Cleo’s shoes?
Ceteris paribus, each of the sentences in (21) presupposes that somebody stole
some of Cleo’s shoes (in the first case this is an entailment, as well), and accord-
170
Chapter 8
ing to my intuitions they will also tend to imply that the person in question
didn’t steal all of Cleo’s shoes. These are admittedly concocted examples, but
real data are no different. In (22), I present three sentences I found on the
internet that involve, respectively, an it-cleft construction, a factive verb, and
a definite description:
(22) a.
It wasn’t me who got some of the font sizes wrong on the Vacancy
Pages, though!
; Somebody got some but not all of the font sizes wrong.
b. I didn’t realize that some of the early church fathers and even the
great reformers [...] believed in the perpetual virginity of Mary.
; Not all the early church fathers believed in the perpetual virginity
of Mary.
c. After all, this was the latest film from Quentin Tarantino, the man
who brought some of the greatest films of the last decade.
; Tarantino didn’t bring all the greatest films of the last decade.
I would like to suggest, then, that presuppositions can give rise to implicatures. More precisely, if a sentence comes with a presupposition ϕ, implicatures
may be derived from ϕ just as if it was asserted. To show that this is actually
a plausible idea, let’s have a closer look at (22a), assuming that it is uttered
in a context in which it is not yet part of the common ground that somebody
got the font sizes wrong. Then the sentence brings two new messages: (i ) that
somebody got the font sizes wrong and (ii ) that that somebody wasn’t the
speaker. (i ) is presupposed and (ii ) is asserted, which means that they don’t
have the same discourse status, but since in both cases the speaker brings new
information to the hearer’s attention, the hearer is entitled to ask himself in
both cases how this information is supposed to contribute to the goals of the
discourse, whether it is relevant, sufficiently specific, and so on. Briefly, in
both cases the hearer is entitled to apply the kind of reasoning that may yield
conversational implicatures.
My proposal requires that presuppositional devices may be used for presenting new information, and since prima facie this requirement is at odds
with the widespread view that presuppositional devices are supposed to carry
given information, I will discuss it a bit further. To begin with, it should be
noted that on some accounts, the notion that presuppositions may carry new
information is entirely straightforward. For instance, one view which recently
has gained some following is that, in general, presuppositions are just parts of
information that are less central to the speaker’s concern than what he wants
8.3 Factives and other presupposition inducers
171
to assert, question, and so on (e.g., Geurts 1999b, Abbott 2000, Simons 2006).
Clearly, if one takes this line, presuppositions may carry new information, and
thus give rise to implicatures.
Things are not as simple if one adheres to what may still be the received
view on presupposition, which originated in Stalnaker’s work of the early 1970s
(Stalnaker 1973, 1974). For, on this account, to presuppose something is
to present it as given, as part of the common ground between oneself and
one’s audience: a presupposition is “an item of presumed common knowledge”
(Stalnaker 1973: 450). However, according to Stalnaker and his followers, just
as one can present old stories as if they were new, a speaker can present new
information as if it were already part of the common ground. In such an
event, the speaker dispenses new information by pretending that the hearer
already has it. For example, if I come late to a meeting, I may apologise by
saying:
(23) I’m sorry I’m late: my bicycle had a flat tire.
In saying this, I presuppose that I have a bicyle. It may be that no one in
the audience knew this beforehand, but they will let me get away with my
presupposition because, in my country, it is perfectly normal for people to
own (and use) bicycles, and because they appreciate that in order to avoid the
presupposition I should have resorted to a cumbersome formulation along the
lines of:
(24) I’m sorry I’m late: I own a bicycle and it had a flat tire.
Hence, new information can, and often will, be conveyed by way of presupposition:
In such a case, a speaker tells his auditor something in part by pretending
that his auditor already knows it. The pretense need not be an attempt
at deception. It might be tacitly recognized by everyone concerned that
this is what is going on, and recognized that everyone else recognizes it. In
some cases, it is just that it would be indiscreet, or insulting, or tedious, or
unnecessarily blunt, or rhetorically less effective to openly assert a proposition
that one wants to communicate. (Stalnaker 1974: 202)
Now, my main point is simply this. Even on an account that explains presupposition in terms of givenness, it is agreed that, in practice, presupposed
information is often new, and that it may be common knowledge between
speaker and hearer that this is so. More briefly, presuppositions may serve as
speech acts for conveying new information. In this sense, they is no different
172
Chapter 8
from assertions (which is not to deny that presuppositions and assertions are
different in other respects), and there is no reason why conversational implicatures couldn’t be derived from them.
Of course, if presuppositions can give rise to implicatures, we will have to
reconsider the standard Gricean line on the relationship between implicatures
and “what is said”. According to Grice, conversational implicatures derive
from what is said, and this is usually taken to imply that they are based
on the truth-conditional content of a statement. However, we know already
that this isn’t quite right, e.g., because non-declarative and multi-sentence
speech acts can have implicatures, as well (cf. §1.3 and §7.3). Although the
overwhelming majority of examples discussed in this book and elsewhere in
the literature involve individual declarative statements, it is just not true that
a conversational implicature has to be based on a single asserted proposition.
Still, even if there is no consensus on the matter, there is a widely held view
that, in the logical order of events, presuppositions are prior to what is said.
Isn’t this a problem? Not as far as I can see. The reason why a statement can
give rise to conversational implicatures is that it is a linguistic act that serves
to convey the information that such-and-such is the case. And as we have seen,
it is universally agreed that, in some cases at least, presuppositional devices
are pressed into service for doing just that. Presupposition and assertion may
be different in other respects, but in this respect they need not be, and that’s
the only thing that matters as far as implicatures are concerned.
To recapitulate, a speaker may choose to present new information by means
of a single statement, a series of statements, or by way of presupposition.
In all these cases, implicatures may ensue, but what makes the last option
particularly interesting is that presuppositions that are triggered in embedded
positions typically behave as if they weren’t; they “project out” of the scope of
negation, modals, and so on. Therefore, though it may seem as if implicatures
derived from such presuppositions exhibit the same projection behaviour, in
reality, that’s not what happens: it is the presuppositions that project out,
not the implicatures they give rise to.
Hence, the problem with factive verbs observed by Chierchia (2004) is just
an instance of a much broader phenomenon. I have argued that the best way
of accounting for the larger pattern is by adopting a discourse perspective:
what is crucial is that presuppositions may serve to bring new information to
the discourse, and this has nothing to with syntax.
8.4 Indefinites
173
8.4 Indefinites
I’ll start this section with a seemingly innocuous detail that, under magnification, will turn out to be deeply significant. One of Grice’s most famous
examples of conversational implicature is the following (cf. §1.1):
(25) A: I am out of petrol.
B: There is a garage round the corner.
Grice glosses this exchange as follows:
B would be infringing the maxim “Be relevant” unless he thinks, or thinks it
possible, that the garage is open, and has petrol to sell ...
(Grice 1975/1989: 32; emphasis added)
Grice’s gloss suggests that A’s reasoning should be unpacked thus:
i. B has said, “There is a garage round the corner”, so apparently he believes
that there is a garage round the corner.
ii. Could it be that B has reason to believe that the garage in question might
be closed?
iii. Presumably not, because then he would have said so.
iv. Hence, I’m going to assume that, for all B knows, the garage is not closed.
What’s interesting about this is that, in order to reason sensibly about B’s
beliefs, A has to accept that there is a garage round the corner, or more
accurately, that B believes there is such a garage. A must be accepting this,
because the question he poses, (ii ), and the inference he draws, (v ), are about
that very garage. The same holds for other examples discussed by Grice:
Anyone who uses a sentence of the form X is meeting a woman this evening
would normally implicate that the person to be met was someone other than
X’s wife, mother, sister, or perhaps even close platonic friend.
(Grice 1975/1989: 37; emphasis added)
In this case, too, the hearer has to accept that, according to the speaker, X is
meeting a woman this evening, so that he can ask himself whether or not that
woman might be X’s wife, mother, sister, or friend.
To see why this is important, consider what an alternative-based analysis
would yield in Grice’s second scenario:
174
Chapter 8
i. S has said: “X is meeting a woman this evening.”
ii. S could have made a stronger statement, e.g., by saying: “X is meeting
his mother this evening.” Why didn’t he do so?
iii. Presumably, it is because S doesn’t believe that the stronger statement is
true. Hence, S doesn’t believe that X is meeting his mother this evening.
This conclusion is weird: though in all likelihood, S’s statement is not about
X’s mother, what we should like to infer is that, as Grice says, “the person
to be met” isn’t X’s mother, but without precluding the possibility that X is
going to see his mother, perhaps en route to the meeting S has in mind.
What went wrong here? The derivation of the fake implicature started out
from the wrong question, namely:
(26) Why didn’t S say: “X is meeting his mother this evening” ?
The question that should have been asked is, in Grice’s own words: “Why
didn’t S say that the person to be met is X’s mother?” The problem with the
alternative-based approach is that it cannot ask this question, simply because
it would be patently improbable to suppose that “The person to be met is X’s
mother” is a suitable alternative to “X is meeting a woman this evening”.
Fortunately, the intention-based stance advertised in Chapter 6 allows us
to ask the right sort of question:
i. S has said: “X is meeting a woman this evening.”
ii. Could it be that S believes the person to be met is X’s mother?
iii. Presumably not, because then S would have said: “X is meeting his mother
this evening.”
iv. Hence, S doesn’t believe that the person to be met is X’s mother.
Apparently, as van der Sandt (1992) was the first to observe, there has to
be a kind of anaphoric link between the speaker’s statement and the hearer’s
question. When the speaker says, “X is meeting a woman this evening”, he
asserts the proposition that X is meeting a woman this evening, but he also
introduces a discourse referent that subsequent utterances can refer back to
and that, crucially, other interlocutors can ask questions about, be it overtly
or tacitly.
The notion of discourse referent was introduced in a seminal paper by
Karttunen (1971), which inaugurated a truly discourse-oriented approach to
interpretation that, in the meantime, has become more or less standard, but
hasn’t had much of an impact on theories of implicature. The main tenet of
8.4 Indefinites
175
this section is that, when we look closer even at the stock-in-trade examples
of conversational implicature, it is evident that such an approach is called for.
In the following, I will develop this point by showing how an intention-based
treatment of Q-implicature might be implemented in Discourse Representation
Theory (DRT), which was the first theory built on Karttunen’s ideas, and
remains one of the most influential frameworks of discourse interpretation to
date (Kamp 1981, Kamp and Reyle 1993, Geurts 1999a).3
DRT is primarily concerned with the interpretation of discourses, not just
sentences. As a discourse unfolds, hearers incrementally construct a model
of what the speaker is saying and implying. Such models consist of two main
parts: a set of discourse referents, which represent the objects under discussion,
and a set of conditions which encode the information that has accumulated
on these discourse referents. For example, when uttered by speaker S, the
sentence in (27) might give rise to the discourse model in (28):
(27) Cleo is reading a book by Theo.
(28) BelS : x
x is a book by Theo
Cleo is reading x
This representation may be given a truth-conditional meaning exactly as one
should expect. That is, (28) is true iff S believes that there is an x such that
x is a book by Theo and x is being read by Cleo. The discourse referent x
is introduced into the model by the indefinite noun phrase “a book by Theo”.
Other expressions may serve to pick up such discourse referents. To illustrate,
suppose S continues his discourse by saying:
(29) It is about logic.
Since this statement is made in the context represented by (28), the pronoun
may be construed as picking up x, and if it is, the effect of (29) is that the
discourse model is extended:
(30) BelS : x
x is a book by Theo
Cleo is reading x
x is about logic
3.
I should note that the following is an extremely sketchy exposition of a hybrid between
Gricean pragmatics and DRT. There are all sorts of issues, technical as well as conceptual,
that remain to be sorted out, and that I have to leave for future research.
176
Chapter 8
One of the attractions of this approach to interpretation is that it has no
need for propositions; no need, that is, for sentence-sized truth-value carriers.
Consider sentence (29) again. It is obvious enough that, taken on its own,
it is neither true nor false. It may be true or false in a given context, such
as the one represented by (28), but even then we first have to decide that
the pronoun is to be interpreted by the discourse referent x, and once that
decision has been made, whatever (29) means has become inextricably tied up
with the information in (28). The primary unit of meaning, on this view, is
the discourse. In the first instance, it is discourses, not sentences, that have
truth conditions.4
Let us now return to implicatures, and the problem posed by:
(31) Last year, a Dutch sailor showed some symptoms of delirium. (= (4))
This is interpreted along the following lines:
(32) BelS : x
x is a Dutch sailor
x showed some symptoms of delirium
As in example (28), the singular indefinite causes the introduction of a new
discourse referent, x (the plural indefinite will do likewise, of course, but we will
ignore that here). On the harmless though admittedly unrealistic assumption
that the previous context is empty, (32) represents what the speaker said by
uttering (31), and it is on the basis of this interpretation that the hearer
will ask himself whether or not S might believe that x showed all symptoms
associated with delirium, which can be represented like this:
(33)
BelS : x
x is a Dutch sailor
x showed some symptoms of delirium
?BelS : x showed all symptoms of delirium
The key point of this model is that the hearer’s question about S’s beliefs is
about the same discourse referent that was introduced by S’s statement. In
4.
This is not to deny that individual utterances have truth-conditional import. Utterances
serve to increment the discourse, and can be assigned meanings in these terms. For example,
we could define the meaning of (29), as used on this particular occasion, as the difference
between the truth conditions associated with (28) and (30).
8.4 Indefinites
177
this sense, there is an anaphoric dependence between S’s statement and the
hearer’s tacit question. From this point on, the reasoning is of the familiar
sort. If the hearer decides to infer that S doesn’t believe that x showed all
symptoms of delirium, his inference can be represented as follows:
(34)
BelS : x
x is a Dutch sailor
x showed some symptoms of delirium
¬BelS : x showed all symptoms of delirium
And if the hearer accepts that S is competent with respect to the proposition
that x showed all symptoms of delirium, there will be a strong implicature, as
well:
(35) BelS : x
x is a Dutch sailor
x showed some symptoms of delirium
¬(x showed all symptoms of delirium)
According to the model in (35), S believes that there was a Dutch sailor who
showed some though not all the symptoms of delirium. Again, it turns out
that what looks like an upper-bounded construal of “some” achieved in an
embedded position is in reality the result of honest-to-goodness pragmatic
reasoning, though it is essential that we adopt a discourse perspective on this
process. That is to say, we have to assume that, when making inferences about
the speaker’s intentional state, hearers reason in terms of discourse referents.
Which is precisely what Grice did in the passages cited at the beginning of
this section.
I have tried to show how we can handle Q-implicatures that seem to arise
within the scope of indefinite noun phrases, without straying from the basic principles of Gricean pragmatics. The proposed solution to this problem
opened up a much larger issue, for I have argued that the heart of the matter is that a theory of conversational implicature should be discourse-oriented
rather than sentence-oriented. When formulated in such general terms, this
is a truism: it is plain that discourse is central to all Gricean thinking about
interpretation. Nevertheless, in the meantime we have come across several
discourse-related factors that have received little or no attention in the literature, but are critically important to the Gricean theory of conversational
implicatures. First, as we observed already in §1.6, in some cases it is a stretch
178
Chapter 8
of discourse, rather than a single utterance, that licenses an implicature. Secondly, I have argued in the last section that implicatures may derive not only
from what an utterance says, in Grice’s sense of the word, but also from what
it presupposes. Thirdly, as argued in the current section, even when implicatures are derived on the basis of what is said, they generally go beyond the
propositional content of an utterance: the implicit question that sets off the
Gricean train of reasoning is very much like a question the hearer might have
asked explicitly in order to clarify the speaker’s intentions, and such explicit
questions will not just be about what the speaker has said. Rather, such questions will typically accept what the speaker has said, at least in the sense that
when the speaker’s utterance introduced new discourse referents, the hearer’s
question can be about those.
My general point, then, is that Gricean pragmatics has a lot to gain from
joining forces with a theory of discourse interpretation like DRT, though it
bears emphasising that the foregoing discussion has scarcely done more than
sketch the bare outlines of such a joint venture. Working out a full-fledged
theory along these lines is not going to be easy, and will inevitably turn up a
fair number of issues, conceptual as well as technical. But the general direction
is clear enough.
8.5 Contrastive construals and lexical pragmatics
Thus far I have argued, for an assortment of allegedly “embedded” implicatures,
that they can be accounted for in a principled way as Q-implicatures, or at
any rate, as essentially involving Q-implicatures. True, we had to assume that
additional factors were involved in each case, but such auxiliary assumptions
as had to be made could always be motivated on independent grounds. Thus
we are left with a single class of problem cases, a selection of which I repeat
here for ease of reference:
(36) a.
b.
c.
d.
Around here, we don’t like coffee, we love it.
I’d rather have a warm bath than a hot one.
Is a parallelogram sometimes or always a square?
If it’s warm, we’ll lie out in the sun. But if it’s very warm, we’ll
go inside and sit in front of the air-conditioner.
For reasons discussed in §7.1, there is no escaping the fact that, in cases
like these, the truth-conditional meaning of “like”, “warm”, etc. has an upper
8.5 Contrastive construals and lexical pragmatics
179
bound. That is to say, in the context of (36a), “like” entails “not love”, and the
same, mutatis mutandis, for the other examples in (36). But even if we have
to concede that marked uses of scalar expressions cannot always be explained
in terms of Q-implicature, that doesn’t mean they can’t be accommodated
within the general Gricean framework. In this section we will see how that
may be done.
Polysemy and the pragmatics of word meaning One of the first intimations of
Grice’s nascent theory of pragmatics occurs towards the end of his 1957 paper:
Again, in cases where there is doubt, say, about which of two or more things
an utterer intends to convey, we tend to refer to the context (linguistic or otherwise) of the utterance and ask which of the alternatives would be relevant to
other things he is saying or doing, or which intention in a particular situation
would fit in with some purpose he obviously has (e.g., a man who calls for a
“pump” at a fire would not want a bicycle pump). (Grice 1957/1989: 222)
As pointed out by Neale (1992), this passage is interesting not only because it
contains the seeds of the views Grice would go on to develop more fully in his
William James lectures, one decade later, but also because his chosen example
is not a conversational implicature but rather a case of lexical ambiguity.
Although one of the established senses of “pump” is “bicycle pump”, it is evident
that someone who calls for a “pump” at a fire is unlikely to have need for a
bicycle pump, because under the circumstances that wouldn’t be relevant to
his purposes. Hence, in this case, Gricean reasoning is employed not to draw
inferences from what the speaker has said, but rather to determine what it is
he wants to say (cf. §1.6).
Actually, though, it is slightly inaccurate to say that the word “pump” is
ambiguous: it is polysemous. A word is standardly considered to be polysemous if it has several senses that are conceptually related to one another.
If you consult a good dictionary for the meaning of a word, it is practically
inevitable that the search will yield more than one sense. In the vast majority
of cases, these senses will be related: polysemy is the rule rather than the
exception. To illustrate, here are some of the senses of “pump”, as given by
the Oxford English Dictionary:
i. A mechanical device for raising water, commonly consisting of a cylinder or tube in which a piston or plunger is moved up and down, and
usually incorporating a valve to prevent back-flow; (more generally) any
mechanical or electromechanical device for raising or moving fluid or
compressing or evacuating gas.
180
Chapter 8
ii. intr. To work a pump (in early use a ship’s pump); to raise or move
water or other fluid by means of a pump.
iii. trans. To inflate (a pneumatic tyre, etc.) by means of a pump. Freq.
with up.
iv. trans. orig. U.S. To enlarge or strengthen (a muscle, part of the body,
etc.) through weight training or steroid use. Chiefly with up.
v. trans. To fire (shots) at a target; to shoot (bullets, etc.) into a person
or thing, now esp. from a machine gun or automatic firearm. Also intr.
All told, the OED lists 40-odd related senses for “pump”, and this bulk is by
no means exceptional: there are multiple senses for the overwhelming majority
of English words, and it is not unusual for a word to have ten senses or more.
(The longest entry in the latest print edition, AD 1989, is for the verb “set”,
which uses 60,000 words to describe some 430 senses.) These observations belie
the naive view that, as a rule, a word has just a single meaning. Polysemy is
clearly the norm, and this fact has an important consequence: it entails that
hearers routinely have to engage in pragmatic reasoning merely in order to
select the right meaning for practically any word used by the speaker.
As illustrated by the excerpt from the OED, your typical lexical entry is
a cluster of senses, which are usually derived from one primary sense, either
directly or indirectly. In the case of “pump”, the oldest sense is (i ), and it is
fairly obvious how the others derive from it. I will also be clear, I trust, that
the web of senses associated with a word is essentially pragmatic in nature.
The fact that dictionaries try to list, for each word in the language, all the
various meanings the word can be used to convey, should not delude us into
thinking that all these senses are purely conventional. To appreciate this point,
consider the following examples (from Nunberg 1979):
(37) a. He can hit the ball two football fields.
b. He hit a home run two games ago.
c. IBM went up three points last week.
In each of these examples, the lexical meaning of the italicised expression
must be shifted so that it will mesh with the rest of the sentence: “football
field” means “length of a football field”, “game” provides a measure of time,
and “IBM” refers to IBM stocks.5 Obviously, such meaning shifts are based
on conceptual relatedness: if α is amongst the established senses of a given
word, then in principle the same word can be used to express any related
5.
For discussion and many more examples, see Nunberg (1978, 1979), Sag (1981), Bosch
(1984), Bartsch (1987), Carston (2002), Recanati (2003), Wilson and Carston (2007).
8.5 Contrastive construals and lexical pragmatics
181
concept β that is rendered salient by α, in a given context, where “context”
is understood sufficiently broadly to encompass the world knowledge shared
between interlocutors. And if a particular meaning shift occurs often enough,
the word in question will have acquired a new established sense: polysemy is
rooted in the context-dependence of word meaning.
In general, lexical interpretation tends to be viewed as a rather passive
affair, the basic idea being that, in order to determine the meaning of a word,
all the hearer has to do is look it up in his mental dictionary. The foregoing
observations paint a rather different picture. Interpreting a word is very much
an active process, in which the hearer continually has to select between large
numbers of possible meanings and has to construct new meanings on the fly.
The dynamics of this process is Gricean: the logic that underwrites the selection and construction of word meanings is similar to that of conversational
implicatures. To drive home this point, consider how a hearer might deal with
the compound nominal “football fields” in (37a):
i. The only use of “football field” I’ve encountered so far is to refer to (well)
football fields.
ii. This is probably not what the speaker has in mind, because if it were,
(37a) would be gibberish. (cf. Quality)
iii. Still, I suppose that that sense is relevant to the intended interpretation
of “football field”. (cf. Relevance)
iv. (37a) would make sense if “football field” denoted a unit which may be
used for measuring the distance a ball is hit.
v. If this unit is to be related to the standard meaning of “football field”, it
might be either the length or width of a football field.
vi. Since a ball is usually hit along the length of a football field, it seems more
likely that the speaker has the first unit in mind, so that’s what I’m going
to assume.
I’m not claiming that hearers actually go through an argument like this every
time they encounter a word. Not only is this an stylised example, but it
is evident that, in many cases, lexical interpretation is a routinised process.
But I do believe it is plausible to assume that the underlying logic of lexical
interpretation is like this.6
6.
An interesting corollary of the Gricean take on polysemy is that it gives us no reason to
suppose that polysemy should be confined to individual words. If the Gricean view is correct,
polysemy might as well occur in complex expressions. There is plenty of evidence that this
182
Chapter 8
In §1.6, I argued that there are two stages at which Gricean pragmatics
is involved in utterance interpretation. First, Gricean reasoning serves to
derive conversational implicatures from speakers’ utterances; it is this use that
has received by far the most attention in the literature and in this book.
Secondly, but first in the natural order of things, Gricean pragmatics helps to
establish the truth-conditional meaning of an utterance, or “what is said”, to
use Grice’s preferred phrase. If the foregoing considerations are on the right
track, this “pre-propositional” use is rife: interpretation processes are guided
by the Cooperative Principle from the lexical level up.7
Narrowing We have seen that one word meaning may spawn another whenever the two are saliently related. One particular way in which two concepts
may be related is that one is more specific than the other. On my first visit
to the United States, I was somewhat taken aback when a waitress who had
just served me returned to ask, “Are you okay?” My first thought was that I
must be looking sick or lonely (or both), but then I realised that she wanted to
know whether I had everything I needed from her. I had just found out that
the waitress was using “okay” in a sense that was somewhat more specific than
I had been accustomed to so far. Examples like this abound; the following are
due to Wilson and Carston (2007):
(38) a. I’m not drinking tonight.
b. Buying a house is easy if you’ve got money.
c. Either you become a human being or you leave the group.
Though it is not entirely unthinkable that someone who utters (38a) intends
not to drink at all, there is a more natural construal on which the object of
is indeed the case, but as the point isn’t central to our purposes, I’ll make do with one
example:
(i)
There was freshly caught fish in the salad.
In (i), it is presumably not just the noun “fish” that acquires a mass interpretation, because
the adjectival phrase “freshly caught” requires that its head have a count meaning. Therefore,
it must the phrase “freshly caught fish”, rather than any of its parts, that is construed in
the culinary sense. Apparently, polysemy is not restricted to the lexicon (and it follows,
incidentally, that the title of this section is not quite accurate).
7. As discussed in §1.6, Grice himself considered that what is said must be fairly close to
conventional meaning. I don’t believe this is right, but nor do I believe that this assumption
is inevitable. On the contrary, it seems to me that the general Gricean framework naturally
leads one to suppose that word meaning may be every bit as context dependent as other
aspects of interpretation.
8.5 Contrastive construals and lexical pragmatics
183
the verb “drink” is constrained to alcoholic beverages. Note that, on this interpretation, the alcohol constraint must be “visible” to the negation operator,
or else the sentence will have the wrong meaning. Similarly, in (38b), “money”
means something like “considerable means”, and in (38c), “human being” is
restricted to members of the species that are “normal” in some sense or other,
and in these cases, too, the restriction must be part of the truth conditions of
the words in question.
In the sentences in (38), sense narrowing happens quite smoothly, because
world knowledge alone suffices to bias the hearer towards a specific interpretation. In other cases, narrowing is helped by other factors, as well:
(39) Bankers are bankers.
If both occurrences of “bankers” had the same meaning, (39) would be a tautology, which it is not, or at least, not necessarily. One way of bringing an
utterance of (39) in line with the Quantity maxim is by supposing that the
second occurrence of the noun is restricted in its application to individuals exhibiting greed, lack of moral fibre, or what have you. Similarly, (40) becomes
informative when the second occurrence of “drinks” is construed as entailing
binge drinking:
(40) When he drinks, he DRINKS.
In this case, the hearer is steered towards a non-standard interpretation by
the contrastive intonation, as well, which we also observe in:
(41) Cleo didn’t open her handbag: she made an incision on the side.
Since Cleo opened her handbag in surgical fashion, her action arguably doesn’t
count as opening in the context of (41).
Whereas in (39) and (40), narrowing offers a way of avoiding a tautological
interpretation, the following sentences would be contradictory if the juxtaposed
terms both had their standard meanings. Hence, one of them will have to be
construed more narrowly:
(42) a.
b.
c.
d.
Julius isn’t rich: he’s RICH. (courtesy Richard Breheny)
The old count didn’t die: he passed away.
They didn’t have sexual intercourse: they fucked. (Horn 1989)
They didn’t fuck: they had sexual intercourse. (ibid.)
While normally speaking, “pass away” would be truth-conditionally equivalent
to “die”, in (42b) it might connote, e.g., an especially stylish manner of expiring,
and the meaning of “die” would then be narrowed down to exclude passing
184
Chapter 8
away, thus understood. (42c) and (42d) are similar, if coarser.8
It will be clear by now, I trust, where my argument is heading: the examples
in (43) (= (36)) can be accommodated in a general Gricean framework simply
as special instances of narrowing:
(43) a.
b.
c.
d.
Around here, we don’t like coffee, we love it.
I’d rather have a warm bath than a hot one.
Is a parallelogram sometimes or always a square?
If it’s warm, we’ll lie out in the sun. But if it’s very warm, we’ll
go inside and sit in front of the air-conditioner.
When construed with the ordinary sense of “like”, (43a) would be a contradiction, but if its meaning is narrowed down to “like but not love”, the sentence
makes perfect sense. In other words, (43a) is no different from the examples in
(42), and something rather similar goes for (43b), where the standard meaning
of “warm” would yield something at least close to a contradiction (as in “I’d
rather live in France than in Paris”), which is averted by an upper-bounded
reading of the adjective. The remaining examples in (43) can be analysed
along the same lines.
What I’m proposing, then, is that the upper-bounded interpretations observed in (43) are special cases of pragmatic narrowing, which itself is a special
case of the context-dependence of lexical meaning. Conversational implicatures don’t play a role in these examples, but this is not to say that they lie
outside the purview of Gricean pragmatics. Rather, they are merely instances
of quite general phenomenon, which has a natural place within the Gricean
scheme of things.
Last call for conventionalism At this juncture, I would like to return to a
suggestion we left dangling at the end of the last chapter (§7.4), where we
considered the possibility of recasting conventionalism as a theory of marked
interpretations of the sort observed in (43). Now that I have sketched the
outlines of a Gricean account of this phenomenon, I can explain why conventionalism is not a viable alternative.
For starters, unlike the Gricean approach, conventionalist theories don’t
8.
There are alternative ways of interpreting examples (42b-d). These sentences might be used
merely to object against the use of an expression on the grounds that it is impolite, overly
bland, or uncouth. Thus interpreted, these are not instances of narrowing, though there is
a still a shift away from the standard meaning, which might be analysed along the lines of
Carston (1996) and Geurts (1998a).
8.5 Contrastive construals and lexical pragmatics
185
have a unified treatment of narrowing to offer. To explain, let’s have another
look at Wilson and Carston’s examples in (38), repeated here for convenience:
(44) a. I’m not drinking tonight.
b. Buying a house is easy if you’ve got money.
c. Either you become a human being or you leave the group.
Since it is plausible to suppose that “drink alcohol” is one of the established
meanings of the verb “drink”, a lexicalist analysis might be defensible for (44a).
But consider how a syntax-based theory would have to explain the preferred
reading of this sentence:9
O
(45) I’m not [ drinking] tonight.
O
O
In order to make this work, the truth-conditional meaning of “ drinking” will
have to be restricted to alcoholic beverages, and given that
performs its
job by eliminating alternatives, the alternative we need is “drink non-alcoholic
beverages”, so that we will get:
(46) I’m not [drinking and drinking non-alcoholic beverages] tonight.
This is blatantly stipulative. Basically, the procedure is to postulate the alternatives that are needed in order to obtain the right meaning; it is entirely
unclear how these alternatives might be justified on independent grounds.
The other two examples in (44) are problematic for any version of conventionalism, though (44c) is perhaps the clearest case. I take it that “ordinary
human being” is not amongst the established lexical senses of the noun “human
being”. True, it may be that speakers normally have ordinary human beings
in mind when they use this expression, but that doesn’t mean that it is an
established sense. If that were the case, “ordinary N” would an established
sense of any noun N (and the same, mutatis mutandis, for other lexical categories), and presumably it would be the most prominent one, too. But then
the preferred reading of (47a) would be (47b), which isn’t plausible at all.
(47) a. If human beings didn’t have bones, they couldn’t drive cars.
b. If ordinary human beings didn’t have bones, they couldn’t drive cars.
Hence, it doesn’t seem very promising to try and argue that “ordinary human
being” is one of the conventional senses of “human being”. As for syntax-based
versions of conventionalism, they don’t do any better in this case than they
did before, and for the same reason: technically speaking, an -based analysis
O
9.
To simplify the exposition a bit, I assume here that
arguments, too.
O can deal with non-propositional
186
Chapter 8
can be made to work, but only at the expense of stipulating alternatives that
are wildly ad hoc.
Next, let’s have a closer look at Breheny’s example:
(48) Julius isn’t rich: he’s RICH. (= (42a))
I take it that the intended interpretation of this sentence is something like:
“Julius is not just rich, but filthy rich”. Referring to the scales in (49), if
the standard, non-upper-bounded meaning of “rich” is rich0 , then the first
occurrence of “rich” in (48) would be interpreted as rich1 and the second
occurrence as rich2 :
(49)
rich0
◦ .....................
.....................
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
◦
rich1
◦ .....................................
rich2
Hence, both occurrences of “rich” in (48) must have undergone narrowing, but
in different ways: whereas rich1 is derived by imposing an upper bound on
rich0 , rich2 is obtained by shifting rich0 ’s lower bound upwards, in the
direction of greater wealth. Now, syntax-based versions of conventionalism
are custom-made to handle the first kind of the narrowing; the second kind
is simply out of their reach. As for lexicalist theories, they would have to
stipulate that rich1 and rich2 are amongst the lexical meanings of “rich”,
which would be ad hoc. In brief: lexicalist and syntax-based theories equally
fail to provide a principled and unified account of narrowing.
8.6 Conclusion
As always, the important thing is to make the right distinctions. In this chapter, I have argued that, regarding “embedded implicatures”, the key distinction
to be made is that between contrastive and non-contrastive environments. In
a contrastive environment, upper-bounded construals may occur in just about
any embedded position. Such construals, which do have truth-conditional import, I propose to analyse in terms of pragmatic narrowing, but they are the
oddball cases, since they require contrastive stress or at least a contrast made
salient by the context. Apart from contrast-induced narrowings, embedded implicatures are far and few between, they never have truth-conditional import,
and they can always be explained in terms of Q-implicature.
8.6 Conclusion
187
This, then, is how embedded implicatures can be accommodated in a
Gricean framework. The alternative is to embrace conventionalism, which
seeks to capture the same phenomena in non-pragmatic terms. However, conventionalist theories have three main drawbacks. First, they leave explanatory
gaps. In particular, weak implicatures, multi-sentence implicatures, and implicatures associated with presuppositions cannot be accounted for within a
conventionalist framework. As a consequence, conventionalist theories will
have to leave these phenomena to pragmatics, and thus are compelled to carve
up the empirical facts in way that goes against nature. Secondly, conventionalist theories are faced with the problem of overgeneration: they produce too
many readings. This problem is more serious for syntax-based versions of the
theory, but lexicalist versions have it, too. Conventionalists try to solve this
problem by stipulating, e.g., that stronger readings are preferred to weaker
ones, but apart from the fact that it is stipulative and merely mitigates the
problem, this strategy yields false predictions in many cases. The fact of the
matter is that conventionalist theories are designed to predict that, in principle, upper-bounded construals are freely available in all embedded positions,
and this doesn’t seem to be correct. Thirdly, conventionalist theories are inevitably ad hoc: their predictions derive from stipulations about the lexicon
and/or the grammar. Hence, the conventionalist approach is at a methodological disadvantage from the start, and since it can hardly be said to outperform
its Gricean competitor, the choice between the two is a no-brainer.
Embedded implicatures have been a formidable bone of contention in the
literature. In particular, they raise the question of whether a considerable tract
of data that, for some decades, was routinely allocated to pragmatics shouldn’t
be relegated to the lexico-grammatical system instead. I have argued that they
shouldn’t, but in doing so have proposed several amendments to the Gricean
framework, some of which were substantial. Most importantly, I believe that
Grice’s views on “what is said” need to be rethought in a number of respects.
First, in my opinion, there is no escaping the fact that the propositional content
of an utterance is determined in large part by pragmatic processes; Grice’s hope
(1975/1989: 25) that what somebody says is “closely related to the conventional
meaning of the words” he uses is not realistic, as far as I can see. Secondly,
we have to reconsider the relation between conversational implicatures and
what is said, and concede that implicatures are derived not only from the
propositional content of a sentence, but also from its presuppositions and even
from longer stretches of discourse. Thirdly (and this is related to the second
point), we have to get away from the traditional, sentence-centred approach
188
Chapter 8
to interpretation. To be sure, Grice was very much aware that discourses, not
sentences, are the primary objects of analysis in pragmatics, but still I believe
that this insight should be carried much further than he did.
Afterword
Looking back on the preceding chapters, we can discern a number of major
themes: some critical, others more constructive. The critical themes all relate
to Grice’s distinction between conventional and non-conventional aspects of
linguistic and paralinguistic content. The seminal idea underlying the Gricean
programme for pragmatics is that conventional content is to be strictly separated from conversational implicature, and one of the recurrent trends in
the literature is for interpretive phenomena that admit of a perfectly decent
pragmatic explanation to be nonetheless relegated to the domain of linguistic
convention. This trend manifests itself variously, but I suspect it is always
rooted in the same basic intuition, namely, that certain aspects of interpretation just don’t feel like bona fide Gricean inferences of the “petrol station”
variety.
To use the hackneyed example one last time, there appears to be an widely
spread gut feeling that “some” has an upper-bounded meaning which is too
robust to be left to Gricean pragmatics. I have tried to dispel this intuition
in sundry ways: by clarifying the notion of cancellability; by arguing that Qimplicatures are not defaults; by trying to show that the robustness of some Qimplicatures admits of a pragmatic explanation and that so-called “embedded”
Q-implicatures can be accommodated by the Gricean theory. I have even tried
to explain what underlies the conventionalist intuition. But the long and short
of it is that there is simply no need for a conventionalist analysis of (some)
Q-implicatures, because the Gricean account will do just fine, and better than
any of the non-pragmatic alternatives we have mustered.
Turning to the more constructive strands in the narrative, I have outlined
a theory which is broad enough to encompass not only scalar implicatures,
which have been the staple of the implicature industry since the 1970s, but
190
Afterword
also free choice inferences, for example. This widening of the theory’s scope
is not only important in its own right; it also brings to light limitations of the
standard Gricean approach to Q-implicatures, in which alternatives play too
central a role. Not that I would want to suggest that alternatives should be
dismissed altogether. On the contrary, alternatives impose vital constraints
on the derivation of some varieties of Q-implicature. But: (i ) other varieties of
Q-implicature aren’t constrained by the availability of alternatives at all, and
(ii ) if and when alternatives are involved, Gricean reasoning shouldn’t begin
with the question why the speaker failed to produced this or that alternative
statement, as the Standard Recipe has it; rather, it should always begin by
considering the speaker’s actions and intentions.
Which brings me to the last leitmotif that I would like to bring out: the
inferential basis of Q-implicatures, in particular, and Gricean reasoning, in
general. The standard picture is very simple, if not downright simplistic: The
speaker utters a sentence, which expresses a proposition (“what is said”, in
Gricean parlance), and there are alternatives to consider: what the speaker
could have said but chose not to. On the standard view, these are the main
ingredients for deriving implicatures. I have argued that, in point of fact,
things are quite a bit more complex than this. First, as mentioned in the last
paragraph, Gricean reasoning is about the speaker’s intentional state. Not
that there can be any doubt that this was Grice’s intention all long, but oddly
enough, this fundamental point often seems to have been lost in the pragmatic
literature. Secondly, while that literature has concentrated its attention on
conversational implicatures licensed by utterances of sentences, there is really nothing special about sentences, since Gricean reasoning equally applies
above as well as below the sentence level. Thirdly, I have argued that we
should take more seriously the undisputable fact that implicature is a discourse phenomenon, which means, for example, that the hearer takes into
account discourse referents introduced by the speaker and that presuppositions, too, may give rise to implicatures. In short, the standard dichotomy
between conventional content, on the one hand, and conversational inferences,
on the other, is really much too simple.
Suppose, if only for argument’s sake, that the theory of Q-implicatures I
have outlined is on the right track. What, then, remains to be done? A lot,
actually, but let me single out three issues. To begin with, we need more experimental data. To date, the bulk of the experimental studies on Q-implicatures
have confined their attention to two words: “some” and “or”. If it is true, as
I have argued, that introspective evidence on implicatures isn’t very reliable,
191
it will be clear that much needs to be done on this score. Secondly, although
I hope to have shown that it is fruitful to develop a discourse-based theory of
implicature, there is no such theory, as yet; what I have provided is a thumbnail sketch, at best. In my opinion, here lies one of the principal theoretical
challenges for the Gricean approach to Q-implicatures and conversational implicature at large. Thirdly, even if, by its very nature, a Gricean theory of
pragmatics is about speakers’ beliefs and desires, it is not a processing theory.
It studies reasoning at a rather lofty level of abstraction, where memory limitations, the fallibility of human reasoning, and hangovers don’t play a role. I
have argued that Gricean pragmatics provides the best general framework for
developing viable processing theories of pragmatic reasoning, but for the time
being such theories are pies in the sky.
Notation and abbreviations
Entailment and truth-conditional equivalence
• “ϕ ⇒ ψ” means that ϕ entails ψ, which is to say that ϕ implies ψ by virtue
of its truth-conditional meaning. Entailment is primarily a relation between
sentences, but it is defined for smaller expressions, as well. For example,
“bright red” ⇒ “red”, “and” ⇒ “or”, and so on.
• “ϕ iff ψ” means that ϕ is true if ψ is true, and vice versa; or in other words:
ϕ and ψ are truth-conditionally equivalent.
Implication
• “ϕ ; ψ” is used, somewhat loosely, to indicate that ϕ may be construed as
implying ψ. In many cases, my argument will be that ψ is conversationally
implicated by ϕ, but occasionally the nature of the implication is left open.
Hence:
• “ϕ 6; ψ” means that ϕ doesn’t imply ψ, except perhaps under very special
circumstances.
?
• “ϕ ;
ψ” means that it is moot whether or not ϕ ; ψ.
Propositional connectives
• “¬ϕ” means that ϕ is not the case.
• “ϕ ∧ ψ” means that ϕ and ψ are both the case.
• “ϕ ∨ ψ” means that at least one of ϕ and ψ is the case.
193
The propositional connectives are used with their classical meanings, and it is
especially important to note that disjunction is inclusive: if ϕ and ψ are both
true, then “ϕ ∨ ψ” is true, as well.
Epistemic notions
• “BelS (ϕ)” means that S believes that ϕ.
• “PossS (ϕ)” means that S considers it possible that ϕ.
For our purposes, Bel and Poss may be viewed as very basic modal operators,
which are interdefinable:
BelS (ϕ) iff ¬PossS (¬ϕ)
I.e., believing ϕ is tantamount to rejecting the possibility that ϕ is false. Conversely:
PossS (ϕ) iff ¬BelS (¬ϕ)
I.e., to consider it possible that ϕ is tantamount to not believing that ϕ is
false.
Note, finally, that “¬BelS (ϕ)” is not equivalent to “BelS (¬ϕ)”: if S doesn’t
have an opinion as to whether ϕ is true, then “¬BelS (ϕ)” is true whereas
“BelS (¬ϕ)” is false. “BelS (¬ϕ)” entails “¬BelS (ϕ)”, but not vice versa.
Bibliography
Abbott, B. (2000). Presuppositions as nonassertions. Journal of pragmatics 32:
1419–1437.
Altmann, G. and M. Steedman (1988). Interaction with context during human
sentence processing. Cognition 30: 191–238.
Asher, N. and D. Bonevac (2005). Free choice permission is strong permission.
Synthese 145: 303–323.
Atlas, J. and S. C. Levinson (1981). It-clefts, informativeness and logical form: radical pragmatics (revised standard version). In P. Cole (Ed.), Radical pragmatics,
pp. 1–61. New York: Academic Press.
Bach, K. and R. M. Harnish (1979). Linguistic communication and speech acts.
Cambridge, Massachusetts: MIT Press.
Baldwin, D. A. (1991). Infants’ contribution to the achievement of joint reference.
Child development 62: 875–890.
Baldwin, D. A. (1993). Infants’ ability to consult the speaker for clues to word
reference. Journal of child language 20: 395–418.
Baron-Cohen, S., A. Leslie, and U. Frith (1985). Does the autistic child have a
“theory of mind”? Cognition 21: 37–46.
Bartsch, R. (1973). “Negative transportation” gibt es nicht. Linguistische
Berichte 27: 1–7.
Bartsch, R. (1987). Context-dependent interpretations of lexical items. In J. Groenendijk, D. de Jongh, and M. Stokhof (Eds.), Foundations of pragmatics and
lexical semantics, pp. 1–26. Dordrecht: Foris.
Beaver, D. and B. Geurts (2010). Presupposition. Stanford encyclopedia of philosophy.
Bezuidenhout, A. and J. Cutting (2002). Literal meaning, minimal propositions,
and pragmatic processing. Journal of pragmatics 34: 433–456.
Blutner, R. (2000). Some aspects of optimality in natural language interpretation.
Journal of semantics (17): 189–216.
195
Bosch, P. (1984). Lexical learning, context dependence, and metaphor. Indiana University Linguistics Club.
Bott, L. and I. A. Noveck (2004). Some utterances are underinformative: the onset
and time course of scalar inferences. Journal of memory and language 51: 437–
457.
Braine, M. D., D. O’Brien, I. Noveck, M. Samuels, R. Lea, S. Fisch, and Y. Yang
(1995). Predicting intermediate and multiple conclusions in propositional logic
inference problems: further evidence for a mental logic. Journal of experimental
psychology: general 124: 263–292.
Braine, M. D. and B. Rumain (1981). Development of comprehension of “or”: evidence for a sequence of competencies. Journal of experimental child psychology 31: 46–70.
Breheny, R. (2006). Communication and folk psychology. Mind and language 21:
74–107.
Breheny, R., N. Katsos, and J. Williams (2006). Are generalised scalar implicatures generated by default? An on-line investigation into the role of context in
generating pragmatic inferences. Cognition 100: 434–463.
Brown, R. (1958). How shall a thing be called? Psychological review 65: 14–21.
Carreiras, M., A. Garnham, J. Oakhill, and K. Cain (1996). The use of stereotypical
gender information in constructing a mental model: evidence from English and
Spanish. The quarterly journal of experimental psychology 49A: 639–663.
Carston, R. (1996). Metalinguistic negation and echoic use. Journal of pragmatics 25: 309–330.
Carston, R. (2002). Thoughts and utterances: the pragmatics of explicit communication. Oxford: Blackwell.
Chemla, E. (2007). Présuppositions et implicatures scalaires: études formelles et
expérimentales. Doctoral dissertation, EHESS, Paris.
Chevallier, C., I. A. Noveck, T. Nazir, L. Bott, V. Lanzetti, and D. Sperber (2008).
Making disjunctions exclusive. The quarterly journal of experimental psychology 61: 1741–1760.
Chien, A. (2008). Scalar implicature and contrastive explanation. Synthese 161:
47–66.
Chierchia, G. (2004). Scalar implicatures, polarity phenomena and the syntax/pragmatics interface. In A. Belletti (Ed.), Structures and beyond, pp. 39–103.
Oxford University Press.
Chierchia, G. (2006). Broaden your views: implicatures of domain widening and
the “logicality” of language. Linguistic inquiry 37: 535–590.
Chierchia, G., D. Fox, and B. Spector (to appear). The grammatical view of
scalar implicatures and the relationship between semantics and pragmatics. In
C. Maienborn, K. von Heusinger, and P. Portner (Eds.), Handbook of semantics.
196
Bibliography
Mouton de Gruyter.
Cohen, L. J. (1971). The logical particles of natural language. In Y. Bar-Hillel
(Ed.), Pragmatics of natural language, pp. 50–68. Dordrecht: Reidel.
Cohen, L. J. (1977). Can the conversationalist hypothesis be defended? Philosophical studies 31: 81–90.
Cruse, D. (1977). The pragmatics of lexical specificity. Journal of linguistics 13:
153–164.
De Neys, W. and W. Schaeken (2007). When people are more logical under cognitive
load: dual task impact on scalar implicature. Experimental psychology 54: 128–
133.
Dennett, D. C. (1987). The intentional stance. Cambridge, Massachusetts: MIT
Press.
Dummett, M. (1973). Frege: Philosophy of language. London: Duckworth.
Eckardt, R. (2007). Licensing “or”. In U. Sauerland and P. Stateva (Eds.), Presupposition and implicature in compositional semantics, pp. 34–70. Houndmills,
Basingstoke: Palgrave Macmillan.
Evans, J. S. and S. E. Newstead (1980). A study of disjunctive reasoning. Psychological research 41: 373–388.
Evans, J. S., S. E. Newstead, and R. M. Byrne (1993). Human reasoning: the
psychology of deduction. Hove, East Sussex: Lawrence Erlbaum.
Fauconnier, G. (1978). Implication reversal in a natural language. In F. Guenthner
and S. Schmidt (Eds.), Formal semantics and pragmatics of natural language,
pp. 289–301. Dordrecht: Reidel.
Feeney, A., S. Scrafton, A. Duckworth, and S. J. Handley (2004). The story of
“some”: everyday pragmatic inference by children and adults. Canadian journal
of experimental psychology 58: 121–132.
Fox, D. (2007). Free choice and the theory of scalar implicatures. In U. Sauerland
and P. Stateva (Eds.), Presupposition and implicature in compositional semantics, pp. 71–120. Houndmills, Basingstoke: Palgrave Macmillan.
Frege, G. (1892). Über Sinn und Bedeutung. Zeitschrift für Philosophie und
philosophische Kritik 100: 25–50.
Frege, G. (1918). Der Gedanke. Beiträge zur Philosophie des deutschen Idealismus 1: 8–77.
Gazdar, G. (1979). Pragmatics: implicature, presupposition, and logical form. New
York: Academic Press.
Geis, M. L. and A. M. Zwicky (1971). On invited inferences. Linguistic inquiry 2:
561–566.
Geurts, B. (1998a). The mechanisms of denial. Language 74: 274–307.
Geurts, B. (1998b). Scalars. In P. Ludewig and B. Geurts (Eds.), Lexikalische
Semantik aus kognitiver Sicht, pp. 85–117. Tübingen: Gunter Narr.
197
Geurts, B. (1999a). Presuppositions and pronouns. Oxford: Elsevier.
Geurts, B. (1999b). Specifics. In B. Geurts, M. Krifka, and R. van der Sandt
(Eds.), Focus and presupposition in multi-speaker discourse, pp. 99–129. Utrecht:
ESSLLI 99.
Geurts, B. (2003). Reasoning with quantifiers. Cognition 86: 223–251.
Geurts, B. (2005). Entertaining alternatives: disjunctions as modals. Natural language semantics 13: 383–410.
Geurts, B. (2006). Implicature as a discourse phenomenon. In E. Puig-Waldmüller
(Ed.), Proceedings of Sinn und Bedeutung 11, pp. 261–275. Barcelona: Universitat Pompeu Fabra.
Geurts, B. (2009). Scalar implicature and local pragmatics. Mind and language 24:
51–79.
Geurts, B. and N. Pouscoulous (2009). Embedded implicatures?!? Semantics and
pragmatics 2(4): 1–34.
Geurts, B. and F. van der Slik (2005). Monotonicity and processing load. Journal
of semantics 22: 97–117.
Green, G. (1996). Pragmatics and natural language understanding. Mahwah:
Lawrence Erlbaum.
Grice, H. P. (1957). Meaning. Philosophical review 66: 377–388. Reprinted in and
cited from Grice (1989: 213-223).
Grice, H. P. (1975). Logic and conversation. In P. Cole and J. Morgan (Eds.),
Syntax and semantics 3: Speech acts, pp. 41–58. New York: Academic Press.
Reprinted in and cited from Grice (1989: 22-40).
Grice, H. P. (1978). Further notes on logic and conversation. In P. Cole (Ed.),
Syntax and semantics 9: Pragmatics, pp. 113–128. New York: Academic Press.
Reprinted in and cited from Grice (1989: 41-57).
Grice, H. P. (1981). Presupposition and conversational implicature. In P. Cole (Ed.),
Radical pragmatics, pp. 183–198. New York: Academic Press. Reprinted in and
cited from Grice (1989: 269-282).
Grice, H. P. (1989). Studies in the way of words. Cambridge, Massachusetts: Harvard University Press.
Guasti, M. T., G. Chierchia, S. Crain, F. Foppolo, and A. Gualmini (2005). Why
children and adults sometimes (but not always) compute implicatures. Language
and cognitive processes 20: 667–696.
Hagoort, P. and C. M. Brown (1994). Brain responses to lexical ambiguity resolution and parsing. In C. Clifton, L. Frazier, and K. Rayner (Eds.), Perspectives
on sentence processing, pp. 45–81. Hillsdale, New Jersey: Lawrence Erlbaum.
Harman, G. H. (1965). The inference to the best explanation. The philosophical
review 74: 88–95.
Hirschberg, J. (1985). A theory of scalar implicature. Ph.D. thesis, University of
198
Bibliography
Pennsylvania.
Horn, L. R. (1972). On the semantic properties of the logical operators in English.
Ph.D. thesis, University of California at Los Angeles.
Horn, L. R. (1984). Towards a new taxonomy for pragmatic inference: Q-based and
R-based implicature. In D. Schiffrin (Ed.), Meaning, form, and use in context:
linguistic applications, pp. 11–42. Washington: Georgetown University Press.
Horn, L. R. (1985). Metalinguistic negation and pragmatic ambiguity. Language 61:
121–174.
Horn, L. R. (1989). A natural history of negation. Chicago University Press.
Horn, L. R. (1997). All John’s children are as bald as the king of France: existential
import and the geometry of opposition. CLS 33: 155–179.
Horn, L. R. (2000). From if to iff : conditional perfection as pragmatic strengthening. Journal of pragmatics 32: 289–326.
Horn, L. R. (2006). The border wars: a neo-Gricean perspective. In K. von
Heusinger and K. Turner (Eds.), Where semantics meets pragmatics, pp. 21–
48. Amsterdam: Elsevier.
Horn, L. R. (2007). Toward a Fregean pragmatics:
“Voraussetzung”,
“Nebengedanke”, “Andeutung”. In I. Kecskes and L. R. Horn (Eds.), Explorations
in pragmatics: linguistic, cognitive and intercultural aspects, pp. 39–69. Berlin:
Mouton de Gruyter.
Horn, L. R. (2009). WJ-40: Implicature, truth, and meaning. International review
of pragmatics 1: 3–34.
Huang, Y. T. and J. Snedeker (2006). On-line interpretation of scalar quantifiers:
insight into the semantic-pragmatics interface. In Proceedings of the 28th annual
conference of the Cognitive Science Society (CogSci 2006), pp. 351–356.
Hurford, J. (1974). Exclusive or inclusive disjunction. Foundations of language 11:
409–411.
Jennings, R. (1994). The genealogy of disjunction. Oxford University Press.
Johnson, S., V. Slaughter, and S. Carey (1998). Whose gaze will infants follow? the
elicitation of gaze-following in 12-month-olds. Developmental science 1: 233–238.
Just, M. A. and P. A. Carpenter (1980). A theory of reading: from eye fixations to
comprehension. Psychological review 87: 329–354.
Kamp, H. (1973). Free choice permission. Proceedings of the Aristotelian Society 74:
57–74.
Kamp, H. (1979). Semantics versus pragmatics. In F. Guenthner and S. Schmidt
(Eds.), Formal semantics and pragmatics of natural language, pp. 255–287. Dordrecht: Reidel.
Kamp, H. (1981). A theory of truth and semantic representation. In J. Groenendijk,
T. Janssen, and M. Stokhof (Eds.), Formal methods in the study of language, pp.
277–322. Amsterdam: Mathematical Centre Tracts 135.
199
Kamp, H. and U. Reyle (1993). From discourse to logic. Dordrecht: Kluwer.
Karttunen, L. (1971). Discourse referents. Indiana University Linguistics Club.
Reprinted in: McCawley (1976: 363-385).
Katzir, R. (2007). Structurally-defined alternatives. Linguistics and philosophy 30:
669–690.
Klinedinst, N. (2007). Plurals, possibilities, and conjunctive disjunction. UCL working papers in linguistics 19: 261–284.
Kutas, M. and S. Hillyard (1980). Reading senseless sentences: brain potentials
reflect semantic anomaly. Science 207: 203–205.
Landman, F. (1998). Plurals and maximalization. In S. Rothstein (Ed.), Events
and grammar, pp. 237–271. Dordrecht: Kluwer.
Levinson, S. C. (1983). Pragmatics. Cambridge University Press.
Levinson, S. C. (2000). Presumptive meanings. Cambridge, Massachusetts: MIT
Press.
Lipton, P. (2004). Inference to the best explanation. London: Routledge.
Liszkowski, U., M. Carpenter, T. Striano, and M. Tomasello (2006). Twelve-and 18month-olds point to provide information. Journal of cognition and development 7:
173–187.
Liszkowski, U., M. Carpenter, and M. Tomasello (2008). Twelve-month-olds communicate helpfully and appropriately for knowledgeable and ignorant partners.
Cognition 108: 732–739.
Matsumoto, Y. (1995). The conversational condition on Horn scales. Linguistics
and philosophy 18: 21–60.
McCawley, J. D. (Ed.) (1976). Syntax and semantics 7: Notes from the linguistic
underground. New York: Academic Press.
McCawley, J. D. (1991). Contrastive negation and metalinguistic negation. In Papers from the 27th regional meeting of the Chicago Linguistics Society. Part 2:
the parasession on negation, pp. 189–206.
Mill, J. S. (1865). An examination of Sir William Hamilton’s philosophy and of
the principal philosophical questions discussed in his writings (2nd ed.). London:
Longmans, Green and Co.
Milsark, G. (1977). Toward an explanation of certain peculiarities of the existential
construction in English. Linguistic analysis 3: 1–29.
Mitchell, D. and D. Green (1978). The effects of context and content on immediate
processing in reading. The quarterly journal of experimental psychology 30: 609–
636.
Morgan, J. (1978). Two types of convention in indirect speech acts. In P. Cole (Ed.),
Syntax and Semantics 9: Pragmatics, pp. 261–280. New York: Academic Press.
Neale, S. (1992). Paul Grice and the philosophy of language. Linguistics and philosophy 15: 509–559.
200
Bibliography
Nieuwland, M. S. and J. J. A. van Berkum (2006). When peanuts fall in love:
N400 evidence for the power of discourse. Journal of cognitive neuroscience 18:
1098–1111.
Noordman, L. G. (1977). Inferring from language. Doctoral dissertation, University
of Groningen.
Noveck, I. (2001). When children are more logical than adults: experimental investigations of scalar implicature. Cognition 78: 165–188.
Noveck, I. and D. Sperber (2007). The why and how of experimental pragmatics:
the case of “scalar inferences”. In N. Burton-Roberts (Ed.), Pragmatics, pp. 184–
212. Basingstoke: Palgrave Macmillan.
Noveck, I. A. and A. Posada (2003). Characterising the time course of an implicature. Brain and language 85: 203–210.
Nunberg, G. D. (1978). The pragmatics of reference. Indiana University Linguistics
Club.
Nunberg, G. D. (1979). The non-uniqueness of semantic solutions: polysemy. Linguistics and philosophy: 143–184.
Onishi, K. H. and R. Baillargeon (2005). Do 15-month-old infants understand false
beliefs? Science 308: 255–258.
Papafragou, A. and J. Musolino (2003). Scalar implicatures: experiments at the
syntax-semantics interface. Cognition 86: 253–282.
Paris, S. G. (1973). Comprehension of language connectives and propositional logical relationships. Journal of experimental child psychology 16: 278–291.
Peirce, C. S. (1931). Collected papers of Charles Sanders Peirce, Volume 1 (edited
by Charles Hartshorne and Paul Weiss). Cambridge, Massachusetts: Harvard
University Press.
Pijnacker, J., P. Hagoort, J. Buitelaar, J.-P. Teunisse, and B. Geurts (2009). Pragmatic inferences in high-functioning adults with autism and Asperger syndrome.
Journal of autism and developmental disorders 39: 607–618.
Pouscoulous, N. (2006). Processing scalar inferences. Doctoral dissertation, EHESS,
Paris.
Pouscoulous, N., I. Noveck, G. Politzer, and A. Bastide (2007). A developmental
investigation of processing costs in implicature production. Language acquisition 14: 347–376.
Recanati, F. (1989). The pragmatics of what is said. Mind and language 4: 295–329.
Recanati, F. (2003). Embedded implicatures. Philosophical perspectives 17: 299–
332.
Recanati, F. (2004). Literal meaning. Cambridge University Press.
Rosch, E. (1978). Principles of categorization. In E. Rosch and B. Lloyd (Eds.),
Cognition and categorization, pp. 27–48. Lawrence Erlbaum.
Russell, B. (2006). Against grammatical computation of scalar implicatures. Jour-
201
nal of semantics 23: 361–382.
Sag, I. (1981). Formal semantics and extralinguistic content. In P. Cole (Ed.),
Radical pragmatics, pp. 273–294. New York: Academic Press.
Sauerland, U. (2004). Scalar implicatures in complex sentences. Linguistics and
philosophy 27: 367–391.
Schulz, K. (2005). A pragmatic solution for the paradox of free choice permission.
Synthese 147: 343–377.
Schulz, K. and R. van Rooij (2006). Pragmatic meaning and non-monotonic reasoning: the case of exhaustive interpretation. Linguistics and philosophy 29:
205–250.
Searle, J. R. (1969). Speech acts: an essay in the philosophy of language. Cambridge
University Press.
Searle, J. R. (1975). Indirect speech acts. In P. Cole and J. Morgan (Eds.), Syntax
and semantics 3: Speech acts, pp. 59–82. New York: Academic Press.
Sedivy, J. C. (2003). Pragmatic versus form-based accounts of referential contrast:
evidence for effects of informativity expectations. Journal of psycholinguistic research 32: 3–23.
Sedivy, J. C. (2007). Implicature during real time conversation: a view from language processing research. Philosophy compass 2/3: 275–496.
Sedivy, J. C., M. K. Tanenhaus, C. Chambers, and G. N. Carlson (1999). Achieving
incremental semantic interpretation through contextual representation. Cognition 71: 109–147.
Simons, M. (2006). Presupposition without common ground. Unpublished
manuscript, Carnegie Mellon University.
Soames, S. (1982). How presuppositions are inherited: a solution to the projection
problem. Linguistic inquiry 13: 483–545.
Southgate, V., A. Senju, and G. Csibra (2007). Action anticipation through attribution of false belief by 2-year-olds. Psychological science 18: 587–592.
Spector, B. (2006). Aspects de la pragmatique des opérateurs logiques. Doctoral
dissertation, University of Paris VII.
Sperber, D. and D. Wilson (1995). Relevance (2nd edition). Oxford: Blackwell.
Sperber, D. and D. Wilson (2002). Pragmatics, modularity and mind-reading. Mind
and language 17: 3–23.
Stalnaker, R. C. (1973). Presuppositions. Journal of philosophical logic 2: 447–457.
Stalnaker, R. C. (1974). Pragmatic presuppositions. In M. K. Munitz and P. Unger
(Eds.), Semantics and philosophy, pp. 197–214. New York University Press.
Storto, G. and M. K. Tanenhaus (2005). Are scalar implicatures computed online?
In E. Maier, C. Bary, and J. Huitink (Eds.), Proceedings of Sinn und Bedeutung
9, pp. 431–445.
Strawson, P. F. (1952). Introduction to logical theory. London: Methuen.
202
Bibliography
Surian, L., S. Caldi, and D. Sperber (2007). Attribution of beliefs by 13-month-old
infants. Psychological science 18: 580–586.
Thagard, P. and C. Shelley (1997). Abductive reasoning: logic, visual thinking, and
coherence. In M. L. Dalla Chiara, K. Doets, D. Mundici, and J. van Benthem
(Eds.), Logic and scientific methods, pp. 413–427. Dordrecht: Kluwer.
van Berkum, J., D. van den Brink, C. Tesink, M. Kos, and P. Hagoort (2008). The
neural integration of speaker and message. Journal of cognitive neuroscience 20:
580–591.
van der Auwera, J. (1997a). Conditional perfection. In A. Athanasiadou and R. Dirven (Eds.), On conditionals again, pp. 169–190. Amsterdam: Benjamins.
van der Auwera, J. (1997b). Pragmatics in the last quarter century: the case of
conditional perfection. Journal of pragmatics 27: 261–274.
van der Sandt, R. (1992). Presupposition projection as anaphora resolution. Journal
of semantics 9: 333–377.
van der Wouden, T. (1997). Negative contexts: collocation, polarity and multiple
negation. London: Routledge.
van Rooij, R. and K. Schulz (2004). Exhaustive interpretation of complex sentences.
Journal of logic, language and information 13: 491–519.
von Wright, G. H. (1968). An essay in deontic logic and the general theory of action.
Amsterdam: North-Holland.
Wainer, J. (2007). Modeling generalized implicatures using non-monotonic logics.
Journal of logic, language and information 16: 195–216.
Walker, R. C. S. (1975). Conversational implicatures. In S. Blackburn (Ed.), Meaning, reference and necessity, pp. 133–181. Cambridge University Press.
Wilson, D. (1975). Presupposition and non-truth-conditional semantics. London:
Academic Press.
Wilson, D. (2000). Metarepresentation in linguistic communication. In D. Sperber
(Ed.), Metarepresentations: a multidisciplinary perspective, pp. 411–448. Oxford
University Press.
Wilson, D. and R. Carston (2007). A unitary approach to lexical pragmatics: relevance, inference and ad hoc concepts. In N. Burton-Roberts (Ed.), Pragmatics,
pp. 230–259. London: Palgrave Macmillan.
Wilson, D. and D. Sperber (2004). Relevance theory. In L. R. Horn and G. Ward
(Eds.), Handbook of pragmatics, pp. 607–632. Oxford: Blackwell.
Wimmer, H. and J. Perner (1983). Beliefs about beliefs: representation and constraining function of wrong beliefs in young children’s understanding of deception. Cognition 13: 41–68.
Zimmermann, T. E. (2000). Free choice disjunction and epistemic possibility. Natural language semantics 8: 255–290.